A causal discovery framework for digital phenotyping Authors: Ahmed Ibrahim Published: 2026 ABSTRACT AbstractDigital phenotyping, the moment-by-moment quantification of human behavior using data from personal devices and sensors, has shown great promise in predicting mental health outcomes. However, the field is reaching a ’predictive plateau,’ where models, while accurate, are often opaque black boxes that offer limited insight into underlying mechanisms of well-being. This paper proposes a fundamental paradigm shift from predictive classification to structural causal modeling. We introduce a two-stage computational framework that first learns unified daily behavioral embeddings from multimodal sensor data using a CNN-based encoder, and then applies neuro-symbolic causal discovery to infer interpretable directed graphs of behavioral–psychological dynamics. In our evaluation, we observed clear signs of this predictive plateau: even deep embedding models performed only slightly better than chance in stress prediction (best AUC = 0.532). By comparison, the causal approach identified candidate time-lagged associations; for example, lower levels of sleep activity (p<0.01) and reduced mobility (p<0.05) often appeared as preceding indicators of stress episodes. Dimensionality reduction via PCA retained five principal components explaining approximately 85% of the variance, enabling post-hoc interpretation of candidate behavioral components such as “Stationary Social Engagement.” We define these components and their associated edge weights as candidate causal biomarkers: hypothesis-generating indicators of possible lagged behavior–stress relationships, rather than confirmed interventional causal effects.Keywords: Digital phenotyping, Neuro-Symbolic methods, Causal biomarkers, Representation learning, Behavioral modelingSubject terms: Biomarkers, Computational biology and bioinformatics, Neuroscience, Psychology, Psychology Introduction A smartphone in your pocket is more than a communication device; it is a scientific instrument of unprecedented scale1. The proliferation of these personal digital devices, particularly smartphones and wearables, has catalyzed the emergence of digital phenotyping, a field defined as the “moment-by-moment quantification of the individual-level human phenotype in situ”2. This framework draws upon the passive collection of sensor data–such as GPS logs, accelerometer readings, and ambient audio–to build detailed, objective depictions of everyday human activity in real-world contexts3. In psychological-wellbeing research, investigators have traditionally depended on self-report instruments and occasional clinical assessments. This method reverses that paradigm by continuously capturing sensor-derived measures–such as movement, location, and ambient sound–to produce an uninterrupted, empirically grounded narrative of individuals’ everyday routines. Consequently, it holds promise for accelerating diagnostic processes and enhancing the accuracy of longitudinal monitoring2. A wealth of empirical evidence attests to the potential of passively recorded sensor information for mental health applications. Studies have, for instance, leveraged patterns of mobile-device usage and physiological signals to anticipate and track conditions ranging from depression and generalized anxiety to bipolar disorder and schizophrenia4. For example, analysis of GPS-derived locational trajectories has revealed that diminished spatial mobility and prolonged home-stay durations correlate with increased depressive symptomatology5. Moreover, aberrations in circadian patterns–inferred from nocturnal smartphone engagement–have been associated with manic episodes6. Evidence indicates that when people with schizophrenia begin making significantly fewer outgoing phone calls over a sustained period, it often signals an imminent relapse7. Continuously examining these call-frequency patterns as digital behavior markers gives clinicians impartial, real-time data that can strengthen and augment standard diagnostic assessments. By providing unobtrusive, continuous monitoring of behavioral and physiological signals, it can inform personalized intervention strategies and may help to alleviate the growing demands placed on mental health care systems worldwide3. Despite impressive advances in predictive accuracy, modern digital phenotyping systems remain largely opaque: deep learning models and engineered feature vectors excel at risk stratification (e.g., predicting depressive episodes or anxiety severity), but reveal little about the causal mechanisms connecting daily behaviors to mental states8,9. We term this limitation the predictive plateau: a state in which correlation-driven models can forecast outcomes but fail to elucidate the underlying causal structure and system dynamics that give rise to those outcomes. Addressing this limitation calls for a change in perspective. Rather than treating mental-health assessment as a one-off classification task, we aim to generate structured hypotheses about lagged relationships between behavioral patterns and subsequent self-reported well-being. In practice, this means moving beyond the question “What is someone’s current state?” to asking “Which behavioral patterns appear to cause later changes in psychological state, and should therefore be prioritized for further validation?” Central to this paradigm is the introduction of candidate causal-discovery biomarkers: model-derived edge-strength scores that summarize possible lagged behavior–psychology relationships rather than confirmed interventional effects. This paper argues for a fundamental paradigm shift to move the field beyond this plateau: a move from predictive classification to dynamic causal modeling. The goal is to evolve from merely forecasting outcomes to constructing interpretable, testable hypotheses about the causal structure of behavioral and psychological dynamics. The combination of ubiquitous computing, advanced machine learning, and time-series analysis in this work addresses a demanding computational challenge and forges a research path with the potential for practical, real-world applications. Specifically, we advance a two-stage computational framework: (1) Unified Behavioral Embedding: A hybrid CNN-Transformer encoder processes raw, multimodal sensor data to learn a semantically grounded, dense latent representation of daily behavior. These latent factors are prealigned with interpretable dimensions (e.g., SleepQuality, SocialActivity) via disentangled or weakly supervised objectives. (2) Neuro-symbolic Causal Discovery: On this learned time-series, we apply advanced causal discovery algorithms (e.g., Neural Granger, NOTEARS) integrated with symbolic reasoning to infer an explicit directed graph of time-lagged causal influences between behavioral embeddings and self-reported psychological states. This two-stage approach leverages the representational power of deep learning while ensuring that the final output–a personalized causal graph–is transparent and testable. Such mechanistic insight opens the door to truly pre-emptive, personalized interventions and the operationalization of causal biomarkers for tracking mental health. To conclude, our paper makes three main contributions: Characterizing the predictive plateau. We critically appraise today’s digital phenotyping workflows–which largely depend on manually crafted variables and correlation-based modeling–and show how they fall short of uncovering the underlying mechanisms that drive behavioral change. Characterizing the predictive plateau. We critically appraise today’s digital phenotyping workflows–which largely depend on manually crafted variables and correlation-based modeling–and show how they fall short of uncovering the underlying mechanisms that drive behavioral change. Causal representation pipeline. We utilize a CNN-based encoder with temporal pooling to learn semantically grounded daily behavioral embeddings. While feature learning is established in mobile sensing, our contribution lies in integrating these representations into a two-stage causal discovery pipeline to move beyond correlational benchmarks. Causal representation pipeline. We utilize a CNN-based encoder with temporal pooling to learn semantically grounded daily behavioral embeddings. While feature learning is established in mobile sensing, our contribution lies in integrating these representations into a two-stage causal discovery pipeline to move beyond correlational benchmarks. Neuro-symbolic causal discovery. We develop a flexible causal discovery pipeline–applicable to both learned and engineered features–that uses neural and symbolic methods (e.g., Neural Granger, NOTEARS) to infer explicit, interpretable causal graphs and operationalize individual candidate causal biomarkers. Neuro-symbolic causal discovery. We develop a flexible causal discovery pipeline–applicable to both learned and engineered features–that uses neural and symbolic methods (e.g., Neural Granger, NOTEARS) to infer explicit, interpretable causal graphs and operationalize individual candidate causal biomarkers. Related work We briefly review three strands of research that inform our proposed paradigm shift: (a) Smart sensing on Digital Phenotyping (b) statistical causal discovery in digital phenotyping, (c) causal representation learning, and (d) deep learning–based causal methods for time series. Smart sensing on digital phenotyping Digital phenotyping is closely connected to the HCI and UbiComp literature on mobile sensing, behavioral inference, and human-centered interpretation of passively collected data. This body of work emphasizes that sensor-derived behavioral features are not neutral clinical measurements10–12; rather, they are shaped by device use, context, missingness, participant routines, and the interaction between people and sensing technologies. These concerns are directly relevant to our framework because causal discovery over passively sensed data depends not only on statistical modeling, but also on how behavioral signals are collected, represented, and interpreted11,13. Recent HCI/UbiComp studies have further highlighted the importance of modeling behavior in ecologically valid settings, designing interpretable sensing systems, and avoiding overconfident claims from noisy behavioral traces11,14. We therefore position our work not as a replacement for this literature, but as a complementary computational framework that converts multimodal sensing data into structured, testable hypotheses about lagged behavior–psychology relationships. Statistical causal discovery in digital phenotyping Early digital phenotyping studies applied classical causal-discovery techniques to handcrafted sensor features. For instance, Jaques et al. applied the PC causal discovery algorithm to daily GPS-derived mobility features (e.g., radius of gyration, number of unique locations visited) alongside self-reported well-being surveys (including stress ratings), uncovering significant time-lagged links whereby days characterized by reduced spatial movement predicted elevated stress levels on the following day15. Lorenz et al. applied linear Granger-causality tests to daily sleep metrics (time in bed and total sleep time) and self-reported depression core-symptom ratings in 22 patients with recurrent major depressive disorder. They found that in 9 of 11 significant cases, variations in sleep metrics Granger-caused changes in next-day depression severity16.. While these works demonstrate the promise of causal inference from passive behavioral data, their reliance on manual feature engineering and linear models limits their capacity to capture nonlinear dynamics and to adapt representations across diverse sensor modalities. Moreover, linear causal methods typically assume stationarity, ignore latent confounders and feedback loops common in behavioral data, and operate on coarse, aggregated features that discard fine-grained temporal dynamics. Techniques such as PCMCI17 and additive noise models18 have been proposed to handle nonlinear and nonstationary time series. To the best of our knowledge, these have generally not been adopted in digital phenotyping contexts. This gap motivates our framework, which unifies end-to-end feature learning with robust, nonlinear causal discovery. Causal representation learning Recent work in causal representation learning aims to uncover latent variables with causal semantics directly from raw data without manual feature design. Methods such as variational autoencoders with causal priors19, disentangled representation frameworks20, and identifiable VAEs under domain shifts21 seek to learn factors of variation that align with assumed causal generative processes. These approaches demonstrate that, under certain structural assumptions, one can recover latent causal variables and their relationships. However, to date, these methods have primarily been applied to vision or synthetic benchmarks; their extension to continuous, multimodal sensor streams in digital phenotyping remains unexplored. Although these representation-learning approaches recover latent factors with provable identifiability under certain assumptions, they typically operate on static data or require strong domain priors to impose causal structure. In particular, methods based on VAEs or disentanglement often ignore temporal dependencies and cannot directly incorporate known behavioral constructs. Moreover, they lack mechanisms to integrate symbolic constraints or human-readable semantics into the learned embeddings, making them ill-suited for personalized, interpretable modeling of time-varying phenomena such as mental-health trajectories. Deep learning based causal methods for time series Recent advances integrate deep learning with causal discovery to handle non-linear, high-dimensional temporal data. Neural Granger Causality replaces linear vector-autoregressive models with neural networks (e.g. MLPs or LSTMs) to detect non-linear time-lagged influences among variables22. NOTEARS formulates graph learning as a differentiable optimization by enforcing a continuous acyclicity constraint, enabling end-to-end learning of sparse causal adjacency matrices23. More recently, Transformer-based approaches such as CausalFormer leverage self-attention to model long-range dependencies and directly output causal adjacency scores from raw time-series inputs24. While these methods demonstrate strong causal recovery on benchmarks like finance and climate, they generally assume homogeneous, single-modality vectors and require well-aligned time steps. They also lack integrated feature-learning mechanisms for heterogeneous sensor modalities and do not produce semantically meaningful embeddings aligned with behavioral constructs. Our work bridges this gap by combining hybrid CNN-Transformer representation learning with neuro-symbolic causal discovery to handle raw, multimodal digital phenotyping streams and output interpretable, individual-level causal graphs. The conceptual leap: from predictive classification to structural causal modeling To fully appreciate the proposed advancement, it is essential to formally articulate the distinction between the prevailing predictive paradigm and the proposed dynamic causal modeling paradigm. This conceptual leap reframes the scientific objective from accurate forecasting to deep, mechanistic understanding, which is a prerequisite for designing the next generation of intelligent and effective digital health interventions. The predictive paradigm: the current State-of-the-Art The standard approach in digital phenotyping for mental health can be characterized as a classification or regression problem. The process begins with the collection of multimodal sensor data and active self-reports over a period of time4. This is followed by a feature engineering step, where a high-dimensional, static feature vector, X, is extracted from the raw data for each user or time window. Subsequently, a machine learning model, f, is trained to learn a mapping f: X → y, where y is a specific clinical outcome, such as a discrete label (’relapse’ vs. ’stable’) or a continuous score (e.g., a predicted GAD-7 anxiety score)5. The primary objective of this entire paradigm is to maximize the predictive accuracy of the model, f(X), on unseen data. While this approach has proven valuable for risk stratification, its core limitation is interpretability. The complex, non-linear models that often achieve the highest accuracy are notoriously opaque. The model may learn that a combination of hundreds of features in X is predictive of depression, but it provides little insight into the structure of the relationships between those features or the temporal dynamics that lead to the outcome. It answers “what” but not “why.” The causal paradigm: a new direction for understanding The structural causal modeling paradigm, in contrast, seeks to uncover the structure of the data-generating process itself. The goal is not simply to predict a label but to infer the system of influences that produces the observed behaviors and psychological states over time. The approach begins by organizing the data into a sequence of time-stamped snapshots, each one capturing the full set of behavioral and psychological measurements at a given moment. This multivariate time series serves as the input for our framework. To manage the practical challenge of sparse psychological reporting, we aggregate behavioral data over a fixed window before applying causal discovery tools. This allows us to identify how consistent behavioral patterns within a period influence subsequent psychological states. We start from the multivariate, time-indexed series where each component of captures a behavioral or psychological measurement at time t. A causal discovery algorithm applied to this series produces a directed graph in which each node v∈ V denotes a specific variable (e.g., sleep quality, social engagement, stress level), and each edge A → B represents a model-derived candidate time-lagged association from variable A to variable B, subject to the assumptions and limitations of observational causal discovery. Accordingly, the central goal of this framework is to derive a model, G, that embodies the system’s temporal behavior and offers a lucid depiction of an individual’s well-being trajectory. Importantly, the causal question addressed in this study is not the estimation of an interventional effect, such as Rather, given observational longitudinal digital phenotyping data, we ask which lagged behavioral components should be retained as plausible causal antecedents of later self-reported stress under assumptions of temporal precedence, observed-variable adjustment, sparsity, and faithfulness. Accordingly, the inferred graph should be interpreted as a structured and testable causal hypothesis, rather than as a confirmed interventional causal model. This shift has important implications. Under the assumptions stated above, the resulting graph provides an explicit and interpretable hypothesis about how behavioral patterns may precede later psychological states. It does not, by itself, establish an interventional mechanism; rather, it organizes observational evidence into a testable model of candidate temporal influences that can guide future validation studies. For example, where the predictive paradigm’s output is a score, the causal paradigm’s output is a potential explanation, such as “A 1-hour decrease in your average sleep duration has been followed by a 20% increase in social isolation 48 hours later.” Such structured hypotheses may help prioritize future validation studies and inform the design of more temporally targeted digital interventions. Table 1 crystallizes the fundamental differences between these two paradigms. Table 1Comparison of predictive and causal modeling paradigms.DimensionPredictive Modeling Paradigm (Current State)Structural Causal Modeling Paradigm (Proposed)Primary GoalMaximize predictive accuracy for a specific outcome.Uncover the underlying dynamic system of influences.Core Question“What is the user’s likely mental state?”“What is the causal structure of the user’s well-being?”Input DataA flat feature vector at a point in time.Aggregated behavioral features from a time-series and corresponding psychological state.Model OutputA class label (e.g., ‘depressed’) or a score.A directed graph representing candidate causal relationships between behavioral patterns and later psychological states.InterpretabilityOften low (e.g., deep learning models).High; the graph explicitly represents hypothesized causal links.Intervention LogicReactive: Trigger intervention when a negative state is predicted.Pre-emptive: Prioritize validation or monitoring when a candidate causal pattern is detected. Comparison of predictive and causal modeling paradigms. A computational framework for dynamic causal phenotyping To realize the vision of dynamic causal modeling, a sophisticated computational framework is required. This section details a novel, two-stage pipeline designed to transform raw, noisy, multimodal sensor data into an interpretable causal graph of behavioral dynamics (we refer to as “candidate causal-discovery biomarkers.”). This framework is explicitly designed to address the key technical challenges of heterogeneity, dimensionality, and temporal complexity inherent in digital phenotyping data. Stage 1: unified behavioral embedding In this stage, we learn semantically grounded behavioral embeddings via a CNN-based encoder with temporal pooling, capturing both local motifs and long-range dependencies across modalities. We begin with raw, minute-level multimodal sensor streams for each modality m, denoted where T is the total number of readings per day (e.g. 1,440 for minute-level data) and each is the sensor value at time t. Our aim is to compress these heterogeneous, high-frequency streams into a sequence of N fixed-length window embeddings where each summarizes the non-overlapping time window (e.g. a 5-minute segment), and N is chosen so that spans a full day. Convolutional Encoder. For each modality m, the raw series is passed through parallel stacks of 1D convolutional filters with kernel sizes 3, 5, 7. These capture local temporal motifs–bursts of physical activity, location transitions, screen-unlock events, etc.–via blocks, each followed by max-pooling to reduce temporal resolution. The local motifs serve as architectural motivations for the encoding process, though they remain to be empirically validated as specific motifs within the latent space. The modality-specific feature maps are then concatenated along the channel dimension into a joint representation. Design Option: Transformer Encoder. In the general framework, the concatenated CNN outputs can optionally be passed through L layers of a standard multi-head Transformer encoder (with H attention heads each). This design choice allows self-attention to capture long-range dependencies–e.g., relating reduced morning social activity to disrupted evening sleep–by attending across all windows. Positional encodings preserve temporal order, and feed-forward sublayers enrich each before layer normalization. However, in the StudentLife experiments reported in this paper, we implemented the CNN+temporal pooling variant described in Section Models implementation. We include the Transformer encoder here as a design option for larger or denser datasets, and as an avenue for future work. Pretraining & Semantic Alignment. To ground each latent dimension in recognizable behaviors, the framework can be trained with a mix of objectives: (1) Masked-modality reconstruction: Randomly mask segments of one or more modalities and task the encoder–decoder to reconstruct them, fostering cross-modal feature sharing. (2) contrastive learning (NTXentLoss) to encourage consistency across modalities. (3) Proxy-label supervision: When auxiliary labels are available (e.g. step-count tiers, screen-time categories), add a small classification head with cross-entropy loss to steer dimensions toward known constructs. (4) Disentanglement regularizer: Apply a Total Correlation penalty on the latent factors to encourage each dimension to capture a distinct source of variation, easing later causal interpretation. In the StudentLife experiments reported in this paper, we implemented only the contrastive NTXentLoss objective. The other objectives are presented here as design options that were not applied in this dataset due to its limited size and sparsity, but may be beneficial in larger or denser cohorts. The output of this stage is a series of daily latent embeddings . Each embedding concisely represents the subject’s behavioral state on day t, integrating measurements of states such as sleep architecture, activity magnitude, and social engagement. These vectors then serve as the input to the neuro-symbolic causal discovery algorithm in the next stage (Fig. 1). Stage 1: Unified Behavioral Embedding pipeline. Raw multimodal streams (GPS, accelerometer, phone logs, wearable signals) are processed via parallel 1D CNNs, concatenated, and passed through a Transformer encoder to produce daily latent embeddings . Transformer block shown as an optional extension; reported experiments used CNN+temporal pooling. Stage 2: neuro symbolic causal discovery Having obtained the daily behavioral embeddings Z from the first stage, these embeddings are analyzed by a neuro-symbolic causal-discovery engine to produce personalized graphs of behavior and psychological states. These graphs are used to define candidate causal-discovery biomarkers, meaning model-derived edge-strength scores that summarize possible causal relationships rather than confirmed interventional effects. Formally, we model candidate causal relationships among the embeddings and parallel self-reported psychological states. where each component denotes the p-th self-reported psychological state at time t (e.g. stress level, mood rating, arousal score, etc.). Our goal is to produce a directed acyclic graph G=(V,E), where consists of one node for each dimension of the behavioral embeddings and each psychological state. Each edge then signifies a model-derived candidate association of variable u at time on variable v at time t, where is the discrete time-lag (in days or windows) over which the candidate causal association is evaluated. For example, an edge indicates that lower inferred sleep quality on day t-1 was selected by the model as a causal candidate antecedent pattern associated with stress on day t. To build this personalized causal graph, we proceed in three steps: first, we train rich neural models to identify candidate time-lagged influences; next, we extract a clean, symbolic directed acyclic graph; and finally, we assign each link a quantitative edge-strength score. Neural Causal Discovery. The process of neural caual discovery involves using advanced time-series models to learn the complex, non-linear dynamics from the data. This extends beyond simple linear models like Granger causality. Two candidate architectures are Neural Granger Causality and NOTEARS. The Neural Granger Causality replaces the linear VAR model with a neural network (such as an MLP or LSTM), allowing it to detect non-linear predictive relationships. Hence, we train a separate feed-forward network or LSTM for each target variable v∈ V to predict its value at time t from the concatenated history . By imposing a sparsity penalty on each model’s input weights, we ensure that only the most influential past variables retain non-zero coefficients, yielding preliminary edge strengths. In the NOTEARS framework, we first form the joint state vector by concatenating the D-dimensional behavioral embedding and the P-dimensional psychological state We then posit a non-linear autoregressive model of the form where is the weighted adjacency matrix we wish to learn–its nonzero entries correspond to candidate causal edges among the |V|=D+P variables; is a noise term capturing unmodeled influences; and denotes the trace (sum of diagonal entries). Enforcing ensures no directed cycles remain in the learned graph. We solve for a sparse adjacency matrix under a continuous acyclicity constraint. The non-zero entries of W directly encode candidate time-lagged causal links among the embedding and state dimensions. Symbolic Graph Extraction. Next, we refine the neural outputs into a sparse candidate graph. We assemble the provisional adjacency matrix from the neural weight or importance scores and then apply three symbolic post-processing steps: first, bootstrap-based significance filtering removes edges that lack statistical robustness; second, for each remaining edge (u→ v), we select the optimal lag by minimizing out-of-sample prediction error; finally, we enforce global acyclicity by solving a minimum feedback-arc set problem, pruning the weakest edges that would induce cycles. This yields a clean, directed acyclic graph G=(V,E) with edges reflecting the strongest, lag-specific influences. Candidate Causal-Discovery Biomarkers. We define each retained edge u → v as a candidate causal-discovery biomarker. Its strength is quantified by the standardized weight from Neural Granger or the normalized entry from NOTEARS. These weights should not be interpreted as interventional effects or estimates of . Instead, they are model-derived edge-strength scores that summarize candidate lagged relationships and can be used to prioritize future confirmatory analyses (Fig. 2). Stage 2: Neuro-Symbolic Causal Discovery. Raw embeddings and self-report streams feed into non-linear causal models (NGC, NOTEARS), symbolic graph extraction, and finally causal biomarker identification. Causal Identification Assumptions and Limitations. The interpretation of the inferred graph depends on assumptions that are difficult to guarantee in observational digital phenotyping data. These include temporal precedence, consistency of measurement, sparsity, faithfulness, and sufficient adjustment for relevant observed variables. In practice, unobserved confounding cannot be ruled out: factors such as personality, baseline mental health, workload, academic deadlines, sleep environment, and social context may jointly influence both behavioral signals and self-reported stress. Therefore, the inferred edges should not be interpreted as identified causal effects. They are best viewed as structured, testable hypotheses about candidate causal relationships that require validation in denser longitudinal datasets or intervention-based studies. A worked example: deconstructing student well-being with the StudentLife dataset This section details the experiments conducted to validate our proposed framework. we first describe our experimental setup, the baseline models used for comparison, and the implementation details of our framework, before presenting and discussing the results. Our evaluation is designed to achieve three primary objectives. First, we aim to quantitatively establish the limitations of the current predictive paradigm by evaluating baseline models that rely on handcrafted features. Second, we demonstrate the end-to-end implementation of our proposed two-stage framework, from learning behavioral embeddings directly from raw data to inferring an interpretable causal graph. Finally, we analyze the components of this graph to showcase its ability to generate novel, data-driven hypotheses about the drivers of well-being. The StudentLife dataset We conduct our evaluation using the StudentLife dataset, which comprises multimodal data from 49 Dartmouth College undergraduates gathered throughout a ten-week academic term. It provides a high-frequency view of student life through passive smartphone sensor streams, including GPS for mobility, accelerometer data for physical activity, and microphone-based features for conversation detection. Crucially, this passive data is contextualized with periodic Ecological Momentary Assessments (EMAs), which provide ground-truth labels for self-reported psychological states such as stress. Owing to its detailed capture of students’ everyday behaviors throughout an academic term marked by fluctuating workloads, this dataset offers an ideal environment for testing and refining well-being models. In assembling our behavioral features and defining outcome variables, we selected a subset of sensor streams from the StudentLife dataset (see Table 2). The passive modalities we employed include data on physical activity, mobility trajectories, social interactions, and patterns of mobile-phone usage. For the ground-truth psychological state, we used the “Stress” EMA survey, which prompted students multiple times per day to rate their current stress level. This combination of high-frequency passive data and periodic active self-reports forms the foundation of our analysis. Table 2StudentLife dataset streams used in this study.Data TypeSensor/SurveyPurpose in Study Passive SensingActivityInferring physical activity states (still, walking, etc.)GPSModeling daily mobility range and location diversityConversationDetecting periods of in-person social interactionPhone LockMeasuring phone engagement and inferring sleepActive ReportingStress EMAProviding the ground-truth target variable for stress StudentLife dataset streams used in this study. Data processing and feature engineering To prepare the dataset for our experiments, we developed two distinct preprocessing workflows–one engineered-feature pipeline for our baseline models and one sequential-input pipeline for our deep-learning framework. The engineered-feature pipeline began by aggregating high-frequency sensor recordings into fixed-length time windows and computing summary statistics (e.g., mean, variance, event counts) within each interval to produce a structured feature matrix for conventional machine-learning algorithms. This process was performed at two levels of granularity to assess the impact of temporal resolution on predictive performance. First, a set of daily handcrafted features was engineered by calculating daily summary statistics (Table 3). These included the total duration of each physical activity state (e.g., still, walking), the cumulative count and duration of conversations, key mobility metrics derived from GPS data such as total distance and location entropy, and phone usage statistics like the daily number of screen unlocks. Second, we applied the same summarization methods to non-overlapping one-hour segments to derive a suite of manually defined features. Consequently, the feature space expanded in dimensionality, enabling the model to capture transient fluctuations in behavior. Table 3Blueprint for Handcrafted Feature Engineering.Behavioral DomainStudentLife Data Source(s)Engineered Features (Daily)Mobility & Activityactivity, gpsDuration of activity states (still, walking, etc.); Total distance traveled; Location entropy; Radius of gyrationSocial InteractionconversationTotal daily duration of conversations; Number of distinct conversationsSleep Patternsphonelock, darkInferred sleep duration (longest inactive period)Device EngagementphonelockTotal number of screen unlocks Blueprint for Handcrafted Feature Engineering. In contrast, the proposed deep learning pipeline was designed to reduce reliance on handcrafted feature construction by learning daily behavioral representations from structured sensor sequences. For this pipeline, the data was organized as sequences of time-stamped sensor events (e.g., GPS coordinates, accelerometer readings, and phone interaction events) for each user-day. These sequences were then provided to the UnifiedBehavioralEncoder, which produced a 64-dimensional representation of daily behavior for downstream causal-discovery analysis. We treat this encoder as an implementation of the representation-learning stage rather than as a standalone architectural contribution. Finally, for all predictive tasks, the target variable was defined as high stress. To account for individual differences in stress reporting, a personalized, binary stress label was created. A user’s self-reported stress for a given time window was labeled as “high” (1) if it exceeded that user’s own average stress level across the entire study period, and “low” (0) otherwise. To create a forecasting task, this binary label was shifted in time. For the daily baseline, the target was defined as the high-stress label of the following day. For the hourly baseline, the target was the high-stress label of the following hour. Models implementation Baseline models: To implement the baseline models, we chose an XGBoost classifier, a powerful and widely-used gradient boosting algorithm, for all baseline experiments due to its high performance on tabular data. For all experiments, the classifier was configured with (the number of boosting rounds) and the evaluation metric was set to logloss. We fixed the parameter to guarantee reproducible results. All other hyperparameters–such as and –were left at the default values provided by the XGBoost Python implementation. All models were trained and evaluated using a rigorous 5-fold group cross-validation scheme, ensuring that all data from a single user was contained within a single fold to prevent data leakage and provide a robust measure of generalizability. We evaluated three distinct baseline configurations. First, we trained a model on the set of daily aggregated features. This represents the most common approach in the literature, where behavior is summarized over a 24-hour period. Second, to test the hypothesis that finer temporal granularity could improve performance, we trained a model on hourly aggregated features, tasked with predicting stress in the following hour. Finally, to understand the predictive power of individual data streams, we conducted a per-modality analysis by training separate daily models on feature subsets corresponding to distinct behavioral domains (e.g., Activity, Social, Sleep). This allowed us to isolate the contribution of each modality to the predictive task. Our framework The first stage of our framework is designed to learn a dense, unified representation of daily behavior directly from raw, multimodal sensor streams. To achieve this, we developed a UnifiedBehavioralEncoder, a neural network implemented in PyTorch. The encoder consists of parallel 1D Convolutional Neural Network (CNN) modules, with one module dedicated to each sensor modality (e.g., Activity, GPS, Phone Lock). Each CNN processes its corresponding time-series data to extract local temporal patterns. The outputs from all modality-specific encoders are then passed through a global average pooling layer and concatenated into a single feature vector. This vector is fed into a final projection head, consisting of two fully-connected layers, which produces the final 64-dimensional behavioral embedding for each user-day. To train the encoder without requiring explicit labels, we employed a self-supervised learning strategy using a contrastive loss (specifically, NTXentLoss). This approach trains the model to generate similar embedding vectors for different augmentations of the same sample, while simultaneously pushing them apart from the embeddings of all other samples in a batch. Training was carried out for ten epochs using the Adam optimizer configured with a learning rate of 1e-3 and a batch size of 16. Upon completion, this stage produces a time-indexed sequence of semantically enriched embedding vectors, which are then supplied to the causal discovery algorithm. The second stage of our framework utilizes the time series of learned behavioral embeddings to infer an interpretable causal graph. Our initial attempts to use a direct 1-day lag (t-1) approach were hindered by the temporal sparsity of the EMA self-reports in the dataset. To create a more robust dataset for causal analysis, we implemented a 3-day look-back window strategy. For each day with a valid stress report, we took the embeddings from the previous three days and averaged them across the 64 dimensions. This gave us with one vector that represented recent behavior in a simpler form, thereby reducing the impact of missing EMAs. The choice of selecting a 3-day look-back window is driven by the goal of reducing the impact of missing EMA responses while maintaining sensitivity to short-term variation. Using a single day as the unit of analysis led to highly unstable estimates because of the frequency of missing self-reports. Longer windows of five to seven days blurred the short-term changes we aimed to capture, since they combined separate stress and recovery periods into broader averages. By comparison, the 3-day window gave us enough stability to handle missing data while still reflecting the day-to-day fluctuations that were most important. To reduce redundancy in the 64-dimensional daily embeddings and improve tractability for causal discovery, we applied principal component analysis (PCA), retaining the top 5 components which together explained over 85%of the variance (The remaining 15% of unexplained variance may carry behavioral signal not captured in our five-component representation). This form of dimensionality reduction preserves the dominant behavioral dynamics while filtering noise, a strategy that has precedent in computational psychiatry and digital phenotyping (e.g25–28.). Consistent with these practices, we retained 5 PCs in order to capture the majority of variance while maintaining interpretability and tractability. Loadings of the retained components were inspected to support semantic interpretation (e.g., PC4 aligning with stationary social engagement). Finally, to infer the causal structure, we explored two neuro-symbolic methods. Our primary causal-structure estimation was performed using the NOTEARS algorithm, configured with a regularization weight of and employing an loss function. In parallel, a Neural Granger Causality model was implemented via a PyTorch-based multi-layer perceptron comprising a single hidden layer of 32 units activated by ReLU. This network was trained for 100 epochs using the Adam optimizer with a learning rate set to 0.01. The final model outputs a directed graph representing the inferred causal relationships (Table 4). Table 4Configuration parameters for the proposed framework.ParameterValue/Setting Stage 1: Unified Behavioral Embedding Model ArchitectureCNN-based EncoderTraining ObjectiveContrastive Loss (NTXentLoss)Embedding Dimension 64 OptimizerAdamLearning Rate 1e-3 Epochs 10 Batch Size 16 Stage 2: Causal Discovery Data Preparation3-day Look-back WindowDimensionality ReductionPCANumber of Components 5 Causal AlgorithmNOTEARSNOTEARS Regularization 0.1 Validation AlgorithmNeural Granger (MLP) Neural Granger Hidden Layer 1 layer, 32 neurons Neural Granger Epochs 100 Configuration parameters for the proposed framework. Evaluation: metrics and results Model performance was assessed using metrics tailored to each task of prediction and causal inference. For the predictive models, we computed the area under the receiver operating characteristic curve (AUC) and the score, obtaining these values through five-fold group cross-validation to ensure a robust evaluation of predictive accuracy. For our causal framework, the primary evaluation is the interpretability and qualitative plausibility of the generated causal graph. We first established a series of baseline models using handcrafted features aggregated at both daily and hourly intervals. The results of our baseline experiments were definitive. The primary model using daily aggregated features achieved an average Area Under the Curve (AUC) of 0.497 and an F1-Score of 0.438, a performance statistically indistinguishable from random chance. We observed that increasing the temporal granularity did not yield better results; the model using hourly features performed similarly, with an AUC of 0.482. Furthermore, our per-modality analysis revealed that no single stream of behavioral data contained a sufficient signal for prediction on its own, with AUC scores ranging from 0.489 for Sleep-related features to 0.499 for Phone Usage (Table 5). Table 5Baseline model performance (5-Fold group cross validation).Model/ModalityAvg. AUC ScoreAvg. F1 ScoreAll Features (Daily)0.4970.438All Features (Hourly)0.4820.434Per-Modality Analysis (Daily Features)Phone Usage0.4990.140Social0.4980.449Activity0.4930.308Sleep0.4890.245 Baseline model performance (5-Fold group cross validation). To further interpret our baseline models, we examined the average importance of each handcrafted feature and plotted the top five in Fig. 3. Here, the daily activity metrics dominate: activity_2 has the highest importance (0.16), followed closely by activity_1 (0.15), activity_3 (0.14), activity_0 (0.14), and sleep_duration_inferred (0.14). Although these features capture meaningful variance, their highest scores remain below 0.16, underscoring why the overall predictive performance hovered around chance. Top five handcrafted feature importances from the baseline model. Bars show the average importance of daily activity and sleep metrics, highlighting that activity_2, activity_1, activity_3, activity_0, and sleep_duration_inferred are the most influential features. To test if a more sophisticated feature representation could overcome this limitation, we evaluated the performance of our deep learning-based embeddings from Stage 1. We tested the 64-dimensional embeddings using both a simple linear model (Logistic Regression) and a powerful non-linear model (XGBoost). The linear model achieved an AUC of 0.482, while the non-linear XGBoost model achieved the highest score of all predictive experiments at 0.532. Despite this slight improvement, the performance remains close to random chance. These results provide empirical evidence of limited predictive performance in the present StudentLife setting and motivate the use of complementary modeling strategies. The primary goal was to move beyond mere prediction to produce an interpretable model of behavioral influences. By applying a 3-day look-back window and the NOTEARS algorithm to our learned embeddings, we obtained the directed causal graph shown in Fig. 4. In this graph, nodes denote the daily stress level and the five principal Behavioral Components (–), and arrow thickness encodes the normalized NOTEARS weight (scaled to [0.5,1.0]). The normalized weight of the pathway is the highest at 1.00. Subsequent pathways– , , , and –exhibit weights of 0.87, 0.69, 0.66, and 0.50, respectively. Under the main NOTEARS configuration, BC4 → Stress received the largest normalized edge weight, followed by BC0, BC3, BC2, and BC1. However, as shown in Section Robustness and sensitivity analysis, stress-directed edges were not stable across bootstrap resampling, PCA dimensionality, or look-back window settings. We therefore interpret BC4 → Stress as an illustrative candidate relationship produced by the main configuration, rather than as a confirmed causal pathway. NOTEARS graph under the main experimental configuration. Nodes represent daily Stress and the five principal behavioral components derived from the preceding three-day window. Edge widths and reported values are normalized for visual comparability only; the underlying NOTEARS coefficients are small, model-dependent quantities and should not be interpreted as effect sizes or interventional causal effects. We applied our self-implemented Neural Granger causality network to the stress levels and behavioral embeddings (Fig. 5). Nodes denote the previous-day stress level and the five principal Behavioral Components (BC0–BC4), and arrow thickness encodes the normalized Granger weight. shows the strongest influence (normalized weight = 1.00), followed by (0.71), (0.69), (0.58), (0.51), and (0.50). The Neural Granger model produced a different ranking of candidate stress-directed relationships, with BC1 → Stress receiving the largest normalized weight. This difference reinforces the model-dependent nature of the inferred edges and supports our interpretation of the graph as a hypothesis-generating output rather than a definitive causal structure. The inferred causal graph generated by the Neural Granger Causality algorithm. Nodes represent daily Stress and the 5 Principal Behavioral Components (PBCs) derived from the preceding 3-day window. Edge widths and reported values are normalized for visual comparability only; the underlying Granger coefficients are small, model-dependent quantities and should not be interpreted as effect sizes or interventional causal effects. Robustness and sensitivity analysis To assess whether the stress-directed edges observed in the main NOTEARS configuration were stable, we performed a bootstrap edge-stability analysis using B=100 resamples under the same five-component PCA representation and three-day look-back window. Figure 6 summarizes the resulting stability pattern. The strongest recurrent structures were not the behavioral-component-to-Stress edges, but rather dependency pairs among the behavioral components themselves. In particular, the BC0/BC1 and BC3/BC4 pairs appeared with high bootstrap inclusion frequencies, suggesting that the learned latent behavioral space contains stable internal interdependencies. By contrast, edges directed toward Stress remained below strong stability thresholds (<50% bootstrap inclusion). We therefore interpret the stress-directed edges reported in the main NOTEARS and Neural Granger configurations as tentative, model-derived candidate relationships rather than stable causal pathways. This finding also helps explain why a purely star-like interpretation centered on Stress would be too simplistic: the most stable structure in the empirical data appears to lie among behavioral components, whereas the mapping from those components to self-reported stress remains more uncertain. Bootstrap stability summary of NOTEARS-inferred behavioral and stress-related relationships in the StudentLife dataset (N=49, three-day look-back window, B=100 bootstrap resamples). Blue nodes denote principal behavioral components and the red node denotes Stress. Solid green arrows highlight recurrent dependency pairs among behavioral components that appeared with high bootstrap inclusion frequency in at least one direction. Dashed red edges denote tentative behavioral-component-to-Stress links that did not reach strong stability thresholds. The figure emphasizes that the most stable empirical structure lies among behavioral components, whereas stress-directed relationships remain configuration-dependent and hypothesis-generating. We also evaluated sensitivity to PCA dimensionality and temporal aggregation. When the number of retained components was varied across , no stress-directed edge appeared consistently across all PCA settings. Similarly, when the look-back window was varied across days, no stress-directed edge appeared in all three settings. These results indicate that the identity of stress-directed candidate relationships is sensitive to both dimensionality reduction and temporal aggregation. Accordingly, the five-component, three-day configuration should be interpreted as the main illustrative configuration used for analysis, rather than as a uniquely validated causal setting. The robustness analysis supports a cautious interpretation of the empirical causal-discovery results. No edge pointing to Stress reached the 80% bootstrap stability threshold, and no stress-directed edge was consistent across all PCA settings or all look-back windows. These findings indicate that the stress-directed edges are sensitive to preprocessing and temporal aggregation choices. We therefore interpret the stress-related edges as tentative candidate relationships rather than stable causal findings. Having established that the stress-directed edges should be interpreted cautiously, we next examined whether the latent behavioral components could still be made human-interpretable. We focused on BC4 as an illustrative case because it received the largest normalized NOTEARS weight in the main configuration, while recognizing that the bootstrap analysis does not support treating it as a stable causal driver of Stress. As a case study, we analyzed by inspecting its PCA loadings, shown in Fig. 7. The loadings reveal that the top five positive contributors to are emb_22 (0.34), emb_15 (0.31), emb_30 (0.23), emb_11 (0.19), and emb_40 (0.17), while the strongest negative contributors are emb_2 (–0.17), emb_19 (–0.22), emb_46 (–0.24), emb_48 (–0.24), and emb_5 (–0.37). These results provide clear, quantifiable insight into which original embedding dimensions drive the behavioral component most strongly linked to stress. Top ten PCA loadings for the behavioral component . Bars are ordered by absolute loading magnitude. The strongest positive contributors are emb_22 (0.34), emb_15 (0.31), emb_30 (0.23), emb_11 (0.19), and emb_40 (0.17), while the strongest negative contributors are emb_2 (–0.17), emb_19 (–0.22), emb_46 (–0.24), emb_48 (–0.24), and emb_5 (–0.37). To ground this statistical fingerprint in real-world behavior, we identified the specific user-days in the dataset that had the highest scores for this component, as detailed in Table 6. A manual inspection of the raw sensor data for these days revealed a consistent pattern. For example, the top-scoring day (user u59 on 2013-06-03) was characterized by high levels of sedentary activity and a large number of distinct in-person conversations. This evidence led us to form a data-driven hypothesis that represents a behavioral phenotype of “Stationary Social Engagement.” Table 6Top 5 user-days with highest BC4 score.User IDDateBC4 Scoreu592013-06-036.99u002013-04-056.99u592013-04-066.34u592013-04-046.15u572013-06-035.27 Top 5 user-days with highest BC4 score. This process demonstrates the framework’s ability to move beyond opaque features to generate interpretable scientific hypotheses. Under the main NOTEARS configuration, BC4 → Stress was selected as an illustrative candidate causal relationship. In this configuration, days characterized by high levels of stationary social interaction may represent a candidate antecedent pattern associated with later stress. This interpretation is hypothesis-generating and should be validated in larger cohorts with denser stress measurements. Synthetic digital phenotyping experiment To make the idea of a “predictive plateau” more concrete and to motivate our causal analysis, we designed a synthetic daily time series that mimics common digital phenotyping signals. The dataset included four behavioral measures–mobility, sleep, social interaction, and phone use–together with a continuous stress variable observed over T=900 days. In constructing stresst, we combined several lagged inputs–mobility, sleep, social, and phone–with modest coefficients. We also added Gaussian noise, both to prevent trivial prediction and to capture the randomness seen in practice. As a result, the simulated data contained clear lagged causal effects but very limited same-day predictability. We next defined a binary high-stress outcome at the 65th percentile. When we fit a logistic regression on same-day inputs (mobility, sleep, social, phonet), the model produced an AUC ≈ 0.56–essentially no better than random. Granger causality tests with lags 1–2 (Bonferroni corrected) revealed several significant temporal links. Notably, sleep predicted stresst, and mobility showed a similar effect (Table 7)). Table 7Granger causality results in the synthetic digital phenotyping experiment (lags 1–2). Reported are raw p-values and Bonferroni-corrected p-values across variables.CauseLagp-valueBonferroni p-valueMobility10.00480.0384Sleep10.00320.0256Phone10.01210.0968Social10.04500.3600Mobility20.01800.1440Sleep20.02700.2160 Granger causality results in the synthetic digital phenotyping experiment (lags 1–2). Reported are raw p-values and Bonferroni-corrected p-values across variables. This synthetic experiment demonstrates that even when predictive performance is near chance, interpretable causal links can be recovered. This supports our central claim: causal discovery provides insight into underlying behavioral dynamics beyond what predictive classifiers can achieve. Ablation: Noise Sensitivity. To check robustness, we ran the simulation again but doubled the variance of the stress noise. As expected, the classifier’s performance dropped further (AUC ≈ 0.52). Even so, Granger causality still picked up lagged relations with statistically significant p-values. In other words, poor predictive accuracy did not prevent us from recovering a meaningful causal structure. Synthetic Experiment Design Choices. To simulate digital phenotyping dynamics, we selected parameters via sensitivity analyses to balance realism and interpretability. The binary high-stress label was defined at the 65th percentile of the continuous stress distribution, rather than the median (50th percentile), because a balanced split yielded artificially high prevalence of “stress episodes” and weakened ecological validity. At the 65th percentile, roughly 35% of the days were marked as high stress, which is closer to a realistic class frequency. Importantly, this threshold still provided a large enough sample for the statistical analyses we report in Fig. 9.). Effect of stress percentile threshold on class balance and predictive AUC. Increasing the threshold reduces the proportion of “high stress” days, but overly high cutoffs (e.g., 90th percentile) yield too few positive samples. The 65th percentile provides a balanced trade-off, with positives and stable predictive AUC. The causal coefficients were deliberately set small–for instance, mobility, sleep, social, and phone. We did this to make stress only mildly influenced by each behavior rather than dominated by one factor. This reflects the empirical findings that no single modality dominates stress prediction, and maintains near-chance predictive performance in same-day classification. Gaussian noise was drawn from , which was chosen after testing smaller variances () that produced artificially high predictive performance (AUC >0.64). Increasing to reduced same-day predictive AUC to ≈ 0.56, in line with our “predictive plateau” premise, while still allowing recovery of lagged causal effects via Granger analysis (Fig. 8). Effect of noise variance on predictive AUC. Low-noise settings (σ =1,2) artificially inflate predictive performance (). At σ =3, predictive accuracy drops to chance level (), consistent with the “predictive plateau” premise, while still allowing recovery of causal effects. Larger variances further degrade predictability without added interpretive benefit. Sensitivity Analyses. To test whether our threshold choice mattered, we repeated the simulation with cutoffs from the 50th to the 90th percentile. The 65th percentile emerged as a reasonable compromise: it produced roughly 35% positives, which seemed more realistic than labeling half the days as stressful, yet avoided the scarcity of cases at higher cutoffs (Fig. 9). Around this point, predictive AUC also showed the least variability. Similarly, varying the noise variance showed that low-noise simulations (σ =1,2) inflated predictive performance (AUC >0.64), contradicting our “predictive plateau” premise. At σ =3, predictive performance dropped to AUC ≈ 0.56, aligning with empirical observations in real datasets, while still allowing recovery of lagged causal edges (Fig. 8). Larger variances further degraded performance without additional interpretive benefit. Discussion Our empirical analysis uncovers the shortcomings of the prevailing predictive framework in digital phenotyping and positions dynamic causal modeling as a practical substitute, thereby refocusing the objective from mere prediction to the elucidation of underlying mechanisms. The experiments yield three primary insights regarding the predictability of stress, the utility of different causal algorithms, and the interpretability of learned behavioral representations. First, our results provide quantitative evidence of limited predictive performance in stress forecasting within this experimental setting: a small (N = 49), relatively homogeneous cohort with sparse EMA coverage. Larger, denser, or more diverse cohorts may yield substantially different predictive outcomes, and these findings should not be read as establishing a general ceiling for predictive modeling across digital phenotyping. The highest AUC score was 0.532, which is consistent with findings from other studies on the StudentLife dataset, such as Bonafonte et al29., who also reported modest accuracies even with extensive feature engineering. These results suggest that, in the present StudentLife setting, further optimization of marginal predictive accuracy alone may provide limited additional insight into the behavioral dynamics associated with stress. Our results, supported by synthetic validation, show that interpretable causal structures can still be explored even when predictive models provide limited utility. This reframes the role of causal modeling as complementary to prediction: its primary value lies in hypothesis generation and mechanistic insight, rather than in incremental predictive gain. Second, our work highlights that even when prediction is unreliable, it may still be possible to organize behavioral signals into interpretable temporal hypotheses. The two causal-discovery methods explored here, NOTEARS and Neural Granger, produced different rankings of candidate stress-directed relationships. Rather than treating this as a contradiction, we interpret it as evidence that the inferred edges are model-dependent and should be used for hypothesis generation rather than as definitive causal findings. NOTEARS emphasizes sparse structural associations under acyclicity constraints, whereas Neural Granger emphasizes lagged predictive relationships. Together, they provide complementary views of candidate temporal dependencies in the learned embeddings. Our framework is not intended as a benchmark of causal discovery algorithms, but rather as a demonstration of the value of shifting from predictive modeling toward interpretable causal structures in digital phenotyping. We focused on two complementary approaches: NOTEARS, which enforces global acyclicity under a linear assumption, and Neural Granger, which models nonlinear temporal dependencies. These methods highlight different aspects of causality—structural on one hand and temporal on the other—and thus provide complementary insights into the embeddings. More broadly, the inferred graph structure is representation-dependent: since the causal discovery stage operates on learned embeddings followed by PCA compression, graphs produced under alternative representation-learning pipelines or latent decompositions may differ substantially, and no single configuration should be treated as a uniquely valid causal model. Other approaches—such as PCMCI(+), DAG-GNN, GraN-DAG, or TCDF—might also be useful, but they usually need larger and more complete datasets to produce stable results, which makes them less practical for the StudentLife cohort. We view the integration of such baselines as an important direction for future work, particularly in larger-scale studies where statistical power is sufficient to support systematic comparisons. The star-like structure of the stress-directed graph should also be interpreted cautiously. It does not imply that human behavior is conditionally independent or that stress is the only organizing node in the behavioral system. Rather, this structure likely reflects the target-centered design of the present analysis, the sparsity of EMA stress labels, PCA compression, and sparsity regularization in NOTEARS. Richer behavioral–behavioral pathways, including feedback loops such as stress affecting sleep or mobility, may be present but are not recoverable with sufficient confidence from the current data. Finally, and most critically, we demonstrated a systematic methodology for making abstract latent variables interpretable. The “candidate causal-discovery biomarkers” produced by our model, such as BC4, are not opaque features. By analyzing the PCA loadings and cross-referencing with raw sensor data from top-scoring days, we decoded BC4 into a human-understandable behavioral phenotype: “Stationary Social Engagement.” This transforms an abstract link in a graph into a novel and testable scientific hypothesis: extended periods of sedentary but socially-dense activity may be a key driver of stress in this population. It is this final step–the translation of a complex computational output into an actionable insight–that showcases the true value of moving beyond prediction to mechanistic modeling. Our causal graphs are viewed as tools for generating hypotheses rather than as final clinical models. They are based on latent components extracted through PCA, which helped reduce redundancy while still capturing about 85% of the variance. To interpret these components, we examined the loadings and linked them to recognizable patterns–for instance, one component appeared to reflect relatively stationary social engagement. These interpretations remain tentative, since they are post-hoc and have not yet been validated outside this dataset. A helpful next step would be to involve domain experts in reviewing these mappings and to test alternative strategies, such as factor analysis or domain-informed embeddings, to improve interpretability. This study has several limitations. First, the StudentLife cohort is relatively small and fairly homogeneous, since it involves undergraduate students from a single institution. This raises concerns about generalizability. That said, the embedding stage draws on millions of minute-level sensor observations in a self-supervised way, which helps offset the limited sample size at the representation level. It is worth emphasizing that this work is intended as a proof of concept; future work on larger and more varied cohorts will be needed to test whether the findings hold with greater statistical power and broader relevance. Second, we rely on a 3-day look-back window to smooth noisy EMA schedules. While this choice improves stability, it may also blur shorter-term temporal patterns. Finally, although NOTEARS and Neural Granger offer complementary views of causality, both approaches rest on assumptions (e.g., linearity for NOTEARS, stationarity for Granger) that may not always hold in real-world behavioral data. Implications for next-generation digital health interventions Reorienting digital phenotyping from a predictive to a causal framework offers significant benefits for the development and implementation of digital health interventions. By uncovering the causal drivers of behavior, this approach supports the creation of interventions that are more efficacious, tailored to individual needs, and anticipatory–especially within the Just-in-Time Adaptive Intervention (JITAI) paradigm. After applying the two-stage framework, the primary output would be a causal graph. Figure 10 presents a hypothetical causal structure in which irregular sleep patterns emerge as the primary antecedent of elevated next-day stress. This heightened stress, in turn, appears to diminish social engagement–operationalized here as the total time spent in conversation. Such results indicate that strategies aimed at regularizing sleep may produce cascading benefits, both by lowering perceived stress and by forestalling social withdrawal within this student cohort. Hypothetical causal structure inferred by the two-stage framework. Irregular sleep patterns act as the primary antecedent of elevated next-day stress, which in turn reduces social engagement (total time spent in conversation). By generating these graphs on a per-user basis, the framework can move towards truly personalized models of behavior, identifying the specific feedback loops that are most influential for each individual. This level of granularity is a significant step beyond current group-level correlational studies. From reactive triggers to pre-emptive intervention Most contemporary JITAIs operate on a reactive basis. They are designed to trigger an intervention when a system predicts that a negative state is imminent or already occurring30. For example, an app might detect patterns associated with high stress and then prompt the user with a mindfulness exercise31. While helpful, this is akin to offering an umbrella once it has already started raining. The proposed causal framework enables a more powerful, pre-emptive approach. By identifying the reliable antecedents of negative states, interventions can be targeted at the root causes before the negative state manifests. A reactive intervention might say, “Our sensors indicate you are feeling stressed. Would you like to try a breathing exercise?” In contrast, a pre-emptive intervention based on a causal model could offer a fundamentally different kind of support: “Our model has learned that for you, a multi-day pattern of spending significant time indoors is often followed by a period of high stress. It’s a nice day out; how about a short walk?” This pre-emptive logic targets a behavior known to be on the causal pathway to a negative outcome, representing a fundamental shift from mitigation to prevention, a key goal in public health. Deep personalization via causal phenotypes The framework’s ability to generate individual-level causal graphs allows for an unprecedented degree of personalization. It is well-understood that the drivers of mental health are heterogeneous; for one individual, social isolation may be the primary trigger for depressive mood32, while for another, it may be sleep disruption33. A one-size-fits-all intervention is unlikely to be effective. Mapping someone’s behavior as a graph of cause and effect lets us trace how one action sparks another. By sorting those connections according to their impact, clinicians can spot the self-reinforcing loops that keep unhealthy habits in place–and then intervene exactly where it matters most. Targeting these key pathways enables interventions to be finely tuned to an individual’s causal profile, embodying the principles of precision psychiatry. For instance, a digital platform might analyze a user’s causal graph and then order its intervention suggestions by the magnitude of their projected effect on key nodes–ensuring that every recommendation aligns tightly with the individual’s specific psychological profile. The rise of the “candidate causal biomarker” This research introduces the potential for a new class of digital biomarker. Currently, digital biomarkers are typically static features or simple statistics derived from sensor data (e.g., average daily step count)2. These markers show differences between groups but may not capture the dynamic processes within an individual. This methodological framework introduces the concept of a causal biomarker, the quantified strength of a particular candidate causal pathway within an individual’s behavioral network. For example, the parameter - which captures the directed effect of irregular sleep on mood as estimated by our neuro-symbolic algorithm - can serve as a personalized indicator of vulnerability. An individual with a very strong link here is phenotypically different from someone for whom this link is weak, even if their average sleep duration is the same. Clinicians could track how these candidate causal biomarkers change over time or in response to therapy. A successful treatment might be one that is observed to weaken a negative causal link (e.g., the influence of stress on sleep) or strengthen a positive one (e.g., the influence of physical activity on mood). This provides a novel, dynamic, and mechanistic way to conceptualize and measure therapeutic efficacy, moving far beyond pre/post symptom questionnaires. Inherent challenges and future research trajectories While our proposed framework introduces novel insights, we acknowledge several key issues. In the sections that follow, we analyze these issues and propose targeted directions for future research. Philosophical and statistical challenges Extracting causal insights from observational data is inherently challenging. Techniques like Neural Granger Causality detect patterns that unfold over time, but they don’t prove a true cause-and-effect relationship in the experimental or philosophical sense. Rather than presenting our model as definitive evidence of mechanism, we should position it as a hypothesis generator. The generated hypothesis can then be tested targeted experimental studies. Practical data and modeling challenges Mobile-sensing studies in real-world settings inevitably encounter gaps in data collection–whether from device failures, drained batteries, or participants occasionally forgetting to carry or charge their sensors. Data collection interruptions can introduce nonrandom omissions that skew subsequent analyses if left unresolved. To address this, our framework integrates statistically principled imputation techniques–ranging from Gaussian-process modeling of missing values to reconstruction via recurrent neural networks–to restore incomplete observations with high fidelity. Data Objectivity and Bias are also concerns; sensor data is a phenotype of the human-device interaction, not just the human. An accelerometer in a phone left on a desk does not reflect user activity. Models must be robust to this noisy and potentially misleading data. Computational Complexity is another factor, as the deep learning models for both representation and causal discovery are intensive. Finally, the Neuro-Symbolic Integration itself is challenging. A primary difficulty is knowledge acquisition: how to best codify complex clinical knowledge into a symbolic form that the model can use without being overly simplistic and constraining the data-driven neural component. Ethical considerations Because digital phenotyping involves the collection of highly personal behavioral metrics, it gives rise to multiple ethical challenges. Data confidentiality and security are enforced through robust end-to-end encryption, strict role-based access controls, and systematic removal of identifying information; participants receive full disclosure of how their data will be used and provide informed consent beforehand34. Ensuring algorithmic fairness requires that training datasets reflect the full diversity of the target population–when models learn from overly homogeneous samples, they can produce biased outcomes that disadvantage underrepresented groups and undermine both validity and trust35. To safeguard user autonomy, the system presents each recommendation in everyday language and allows individuals to accept, question, or ignore automated suggestions. Nevertheless, surfacing inferred causal links can provoke anxiety or foster undue dependence on the technology. To make this work, we need clear, user-friendly designs paired with ongoing clinician support–so professionals can guide users in interpreting and applying the system’s suggestions responsibly. Future research trajectories The proposed framework serves as a foundation for an extensive research program. Future work could proceed along several exciting trajectories. One direction is the Integration of New Data Modalities, using the modular architecture to incorporate text from social media or physiological signals from advanced wearables like heart rate variability. Another is the Validation on Diverse Datasets, testing the framework’s generalizability on public datasets like Tesserae (workplace well-being) or CrossCheck (schizophrenia). The exploration of Advanced Causal Inference techniques is also a key path, moving beyond predictive causality towards methods based on structural causal models or counterfactual reasoning. Lastly, Integration with Large Language Models (LLMs) presents a promising frontier for extracting features from unstructured text (e.g., journal entries), providing natural language interpretations of the inferred causal graphs, and generating personalized, narrative-based intervention content. Future work should also explore the integration of richer clinical modalities, particularly voice. Recent work has shown that methodological choices in voice data collection are critical to the reliability and clinical validity of depression detection. For example, Vilenchik et al. proposed a structured data collection framework demonstrating that carefully designed, clinically grounded voice acquisition can improve depression detection compared with less controlled collection settings36. Building on this direction, future extensions of our framework could incorporate vocal features as proximal indicators or mediating variables within the causal graph, thereby strengthening the link between digital phenotyping signals and clinically meaningful mental health outcomes. For example, vocal prosody or speech-energy features could be modeled as observed nodes between passive behavioral context and self-reported mood, allowing future work to test whether voice acts as a mediator, proximal outcome indicator, or confounding-adjustment variable. Conclusion: a new trajectory for computational psychiatry This paper addresses the “predictive plateau” within the constraints of the StudentLife dataset and experimental setting, where predictive models may achieve limited accuracy while providing little insight into the causal organization of behavioral and psychological states. Our empirical evaluation on the StudentLife dataset showed this challenge: models using both handcrafted and learned features achieved near-chance performance in daily stress prediction. In response, we introduced a two-stage computational framework that combines learned behavioral embeddings with causal-discovery methods to generate interpretable, testable hypotheses about candidate causal behavior-stress relationships. The primary contribution of this work is not the identification of a confirmed causal effect, but the demonstration of a hypothesis-generating pipeline for exploratory causal discovery in digital phenotyping. In the main configuration, the framework identified a candidate behavioral component that we interpreted post hoc as “Stationary Social Engagement.” However, robustness analyses showed that stress-directed edges were sensitive to bootstrap resampling, PCA dimensionality, and look-back window length. Accordingly, this finding should be interpreted as a model-derived candidate relationship that requires validation in larger and denser longitudinal datasets. Ultimately, the study reframes digital phenotyping from a purely predictive task toward a more interpretable and temporally structured form of behavioral analysis. By treating inferred edges as candidate causal-discovery biomarkers rather than confirmed interventional effects, the framework provides a cautious path for generating clinically relevant hypotheses that can be tested in future observational, counterfactual, or intervention-based studies. Author contributions The Author wrote the manuscript, design and implemented the exepriment. Funding This work was funded by the University of Jeddah, Jeddah, Saudi Arabia, under grant No. (UJ-24-DR-1166-1). Therefore, the authors thank the University of Jeddah for its technical and financial support. Data availability The StudentLife dataset used in this study is publicly available: https://www.kaggle.com/datasets/dartweichen/student-life To facilitate reproducibility, we will provide our preprocessing scripts, configuration files, random seeds, and group cross-validation splits after publication upon request. This will enable replication of both predictive baselines and causal discovery results. Declarations Competing interests The authors declare no competing interests. Footnotes Publisher’s note Springer Nature remains neutral with regard to jurisdictional claims in published maps and institutional affiliations. References Associated Data Data Availability Statement The StudentLife dataset used in this study is publicly available: https://www.kaggle.com/datasets/dartweichen/student-life To facilitate reproducibility, we will provide our preprocessing scripts, configuration files, random seeds, and group cross-validation splits after publication upon request. This will enable replication of both predictive baselines and causal discovery results.