PULSE: Agentic Investigation with Passive Sensing for Proactive Intervention in Cancer Survivorship Journal: IMWUT Volume: 0 0 0 9 CCS: Human-centered computing Ubiquitous and mobile computing systems and tools CCS: Computing methodologies Natural language processing
Zhiyuan Wang, Ariful Islam, Indrajeet Ghosh, Xinyu Chen, Katharine E. Daniel, Subigya Nepal, Philip Chow, Laura E. Barnes · 2026
Read PULSE: Agentic Investigation with Passive Sensing for Proactive Intervention in Cancer Survivorship Journal: IMWUT Volume: 0 0 0 9 CCS: Human-centered computing Ubiquitous and mobile computing systems and tools CCS: Computing methodologies Natural language processing by Zhiyuan Wang, Ariful Islam, Indrajeet Ghosh, Xinyu Chen, Katharine E. Daniel, Subigya Nepal, Philip Chow, Laura E. Barnes (2026) in the Tower of Babel, a free online library of public-domain works preserved in clean, searchable text. No sign-up. The full text is served as a plain-text file at a predictable URL, where text is present: https://antarctic-labs.com/tower-of-babel/arxiv-2605-17679.txt.
Introduction
Zhiyuan Wang and seven colleagues built PULSE, a system that helps cancer survivors get emotional support even when they have not filled in a diary entry. Language-model agents pick from eight tools to read phone sensing data, compare behavior with personal baselines, and pull up similar past cases. In tests on 50 survivors, the agent-led approach reached balanced accuracy near 0.74 for spotting a desire for emotion regulation, beating fixed-summary methods.
Catalog record
| CREATOR | Zhiyuan Wang, Ariful Islam, Indrajeet Ghosh, Xinyu Chen, Katharine E. Daniel, Subigya Nepal, Philip Chow, Laura E. Barnes |
|---|---|
| YEAR | 2026 |
| COLLECTION | PAPERS |
| FORMAT | TXT |
| FILE SIZE | 94794 |
| LICENSE | http://creativecommons.org/licenses/by-sa/4.0/ |
| RIGHTS | CC BY SA |
| SOURCE | arXiv |
Source: https://arxiv.org/abs/2605.17679
Browse the full library · About the Tower of Babel
Full text
The opening of the work, from the library's hosted text:
PULSE: Agentic Investigation with Passive Sensing for Proactive Intervention in Cancer Survivorship Journal: IMWUT Volume: 0 0 0 9 CCS: Human-centered computing Ubiquitous and mobile computing systems and tools CCS: Computing methodologies Natural language processing Zhiyuan Wang, Ariful Islam, Indrajeet Ghosh, Xinyu Chen, Katharine E. Daniel, Subigya Nepal, Philip Chow, Laura E. Barnes Abstract . Cancer survivors face elevated rates of depression, anxiety, and general emotional distress, yet the precise moments they most need support are often the moments when self-report is sparse, a phenomenon we term the diary paradox . Passive smartphone sensing offers a continuous, unobtrusive alternative, but prior sensing-based affect prediction has been limited by an accuracy ceiling, suggesting a bottleneck not only in available data, but in how behavioral signals are interpreted. We present Pulse , a system that shifts from fixed feature pipelines to agentic sensing investigation : LLM agents equipped with eight purpose-built tools autonomously query smartphone sensing data, compare current behavior against personalized baselines, and calibrate inferences through retrieval-augmented population-level comparisons. Rather than receiving pre-formatted feature summaries, agents decide which modalities to inspect, how far back to look, and how deeply to investigate, mirroring hypothesis-driven clinical reasoning. We evaluate Pulse through a 2 × \times 2 factorial design crossing reasoning architecture (structured vs. agentic) with data modality (sensing-only vs. with diary) on 50 cancer survivors from a longitudinal study of cancer survivors. Agentic reasoning is the primary driver of performance: agentic multimodal agent achieves balanced accuracy of 0.743 for emotion regulation desire with diary and sensing data, while agentic agents predict intervention availability at 0.713 with passive sensing data only. These results suggest that agentic investigation may be a cornerstone for unlocking the clinical value of passive sensing, advancing the feasibility of proactive just-in-time mental health support. Figure 1 . Pulse addresses ambiguity in brief diary entries by integrating passive behavioral sensing with agentic investigation. A diary entry alone admits multiple interpretations (left); always-on smartphone sensors reveal behavioral context (center); the Pulse agent autonomously queries sensor data through purpose-built tools to disambiguate affect and predict intervention need (right). A three-panel figure showing the PULSE pipeline: (1) an ambiguous diary entry with three possible interpretations, (2) passive smartphone sensors including GPS, motion, screen, and diary, and (3) the PULSE agent using tools to produce binary state predictions for positive affect, negative affect, regulation desire, and intervention availability. 1. Introduction More than 18 million cancer survivors live in the United States in 2026, and almost a third experience clinically significant depression, anxiety, or emotional distress in the years following treatment ( 59 ; 39 ) . Clinical guidelines now recommend routine screening and management of these symptoms ( 6 ) , yet timely and scalable mental health interventions remain underdeveloped. Just-in-time adaptive interventions (JITAIs) aim to use mobile technology to offer support to survivors at precisely the moments when support is most needed and most likely to help ( 42 ) . Effective JITAIs could reduce the mental health burden of cancer survivors at scale. However, an effective JITAI must do more than detect distress—it must infer whether support is currently desired and whether the survivor is available to engage with different types of support were it to be offered. In JITAI terminology, this process is determined by tailoring variables , intervention options , decision points , and decision rules , or components that determine when, whether, and how to intervene from real-time signals ( 43 ) . The dominant approach to estimating if an intervention should be delivered relies on ecological momentary assessment (EMA), which are brief self-report surveys delivered multiple times daily to a survivor through their mobile phone ( 57 ) . While EMA provides invaluable ground truth for research, it introduces a fundamental paradox when used as the input to a proactive support system: the very moments when intervention is most needed are precisely when self-report data are most likely to be absent ( 20 ; 41 ) . We call this the diary paradox . For example, one cancer survivor experiencing acute distress may stop responding to surveys because they are too overwhelmed; another survivor caught in a depressive episode may lack the motivation to complete a diary entry. As a result, the data that would be most clinically informative for these hypothetical survivors—those that capture periods of high distress and low functioning—would be missing not at random. In CALLM ( 71 ) , the authors highlighted that diary text is highly predictive of emotional states when available, but it cannot be used to predict emotional states when the diary itself is absent. A truly proactive system must be able to reason about the user’s state from data that requires no active participation. Passive smartphone sensing provides exactly this: continuous, unobtrusive streams of behavioral data—movement patterns, location traces, screen usage, sleep timing, and social communication—that accumulate regardless of the user’s engagement with diaries or other self-report prompts. A decade of mobile sensing research, from StudentLife ( 68 ) and early digital phenotyping ( 55 ; 48 ) to GLOBEM ( 76 ) , has established that these behavioral signals correlate with mental health outcomes. Yet, the translation from correlation to clinically useful prediction has been frustratingly slow. Prior work reveals a persistent prediction ceiling: traditional machine learning models achieve approximately 0.52-0.60 balanced accuracy on passive sensing-based affect prediction, barely above chance for many targets, and this ceiling persists across datasets and feature engineering approaches ( 76 ; 2 ; 34 ) . Why does this ceiling exist? While there are several possibilities, we argue that the problem is not in the data itself but in the reasoning that traditional models apply to these data. Fixed feature pipelines extract the same summary statistics from every user on every day, regardless of context. They cannot notice that a particular user’s 3:00 AM screen session is anomalous relative to their personal baseline, or that a sudden drop in GPS mobility following a week of high activity might signal a withdrawal pattern. They cannot synthesize sleep disruption, reduced social communication, and increased late-night phone use into a coherent clinical narrative. The bottleneck lies not in feature extraction but contextual behavioral signal interpretation ( 44 ) —the kind of integrative reasoning that a clinician performs when interpreting a patient’s behavioral record. Recent advances in large language models (LLMs) suggest a different path forward. LLMs can synthesize behavioral evidence in natural language, and prior work has shown that they can reason about both mental health text and wearable sensing summaries ( 26 ; 77 ; 74 ) . But most systems still hand the model a pre-formatted summary and ask it to reason using whatever the fixed pipeline has already chosen to expose. The model may be eloquent, yet it still cannot decide what to inspect, which anomaly to pursue, or how to compare today’s behavior against the right personal baseline. In this paper, we ask two questions: (1) can LLM agents predict intervention opportunities from passive sensing data, and (2) does giving the LLM autonomy to investigate sensor data—dynamically deciding which modalities to query, how far back to look, and which anomalies to pursue—improve prediction over fixed analytical pipelines? We introduce Pulse , a system