SEANN: A domain-informed neural network for epidemiological insights Authors: Jean-Baptiste Guimbaud, Marc Plantevit, Léa Maître, Rémy Cazabet, Vijay Kumar, Vijay Kumar, Vijay Kumar Published: 2025 ABSTRACT In epidemiology, traditional statistical methods such as logistic regression, linear regression, and other parametric models are commonly employed to investigate associations between predictors and health outcomes. However, non-parametric machine learning techniques, such as deep neural networks (DNNs), coupled with explainable AI (XAI) tools, offer new opportunities for this task. Despite their potential, these methods face challenges due to the limited availability of high-quality, high-quantity data in this field. To address these challenges, we introduce SEANN, a novel approach for informed DNNs that leverages a prevalent form of domain-specific knowledge: Pooled Effect Sizes (PES). PESs are commonly found in published Meta-Analysis studies, in different forms, and represent a quantitative form of a scientific consensus. By integrating PES into the training loss, we demonstrate—under controlled simulations—significant improvements in predictive generalization and in the epidemiological plausibility of the learned relationships, relative to a domain-knowledge agnostic neural network. 1 Introduction Historically in epidemiological studies, the impact of environmental health associations was largely studied using a ‘one-exposure-one-health-effect’ approach [1]. While such targeted approaches are informative, scaling them to the diversity of existing environmental factors is expensive and they can miss complex interactions and unaccounted confounders. The exposome paradigm emerged to address these limitations, proposing to consider the totality of individuals’ environmental exposures. This holistic framework allows investigating the combined effects on health from diverse environmental factors, including urban, chemical, lifestyle, and social hazards. Analyzing such complex mixtures of exposures requires advanced modeling techniques capable of handling high-dimensional data from observational studies [2,3]. While traditional biostatistical methods remain important, recent advancements in machine learning, particularly Deep Neural Networks (DNNs), offer a powerful alternative for discovering complex patterns in high-dimensional data. Their use, however, poses specific challenges. Their need for large and high-quality datasets and their lack of interpretability are significant obstacles, especially in healthcare, where data can be scarce, noisy, and ethically sensitive. In dealing with such data, purely data-driven approaches can lead to unsatisfactory results, such as capturing spurious associations. In addition, purely data-driven methods do not comply with known natural laws (e.g., biological pathways). Incorporating additional knowledge to complement the training data has proven its ability to address those issues in various domains [4]. In epidemiology and other cumulative empirical science, scientific knowledge emerges from a consensus among multiple studies, focusing on the same question but in slightly different settings. In the context of the exposome, in which we search for the relation between variables of interest and a target outcome, each study estimates these relations, and these estimations are aggregated across several studies in meta-analyses [5], to derive a more reliable and statistically robust indicator, namely a Pooled Effect Size (PES) [6]. PESs thus represent a quantitative formulation of a scientific consensus. Considered one of the most reliable forms of information in the field [7], PESs have been used to compute literature-only health risk scores (e.g. [8]) and informed risk scores (e.g., [9]) using traditional biostatistical methods (e.g., logistic and linear regression). However, the integration of PESs within machine learning models, particularly DNNs, has not been explored yet, despite their potential to improve the reliability of these models in scarce and noisy data settings. In this work, we introduce SEANN (Summary Effects Adapted Neural Network), a novel approach designed to integrate prominent forms of PESs, namely Odds Ratios (ORs), Relative Risks (RRs), and Standard Regression Coefficients (SRCs) [10,11] directly into DNNs learning process. By incorporating PESs as a form of scientific prior, SEANN aims to enhance the model’s generalizability and ensure that extracted relationships are scientifically plausible. This is achieved through a custom loss function that penalizes deviations from integrated PESs measured through differences in prediction when perturbing the inputs. SEANN addresses challenges posed by Deep Neural Networks (DNNs) that limit their use in epidemiological contexts, compensating for the limited observational data typically available in exposome-wide studies and improving the generalizability and trustworthiness of computed risk indicators. Additionally, by incorporating knowledge from well-known relationships, SEANN can better characterize those that are less studied. The paper is structured as follows: in Sect 3, we introduce SEANN; In Sect 4, we perform a series of experiments on synthetic scenarios illustrating the method’s benefits in a controlled environment. More specifically, we refer to improved prediction accuracy in noisy contexts and improved reliability of interpretation using XAI. Finally, Sect 5 discusses the significance of those results and concludes. 2 Related work Machine Learning (ML) models and DNNs in particular often rely heavily on both the quantity and the quality of the available training data. In many fields, including healthcare, securing vast and representative datasets poses significant challenges due to ethical, logistical, and technical constraints that would consequently impact the reliability of obtained predictions [12]. Beyond predictive performances, most machine learning procedures do not consider the underlying mechanisms at play (e.g., biological pathways, physical rules, etc) when they learn patterns within the data. As such, they may learn and amplify potential biases, particularly when the data is noisy and incomplete [13]. Informed Machine Learning (IML) addresses these challenges by combining data-driven learning with domain-specific knowledge, leveraging a hybrid approach that can enhance both predictive accuracy and model interpretability [4]. While it is still an emergent stream of research, the number of papers published per year in healthcare alone has approximately doubled every year, starting from 10 published studies in 2018 to 58 in 2021, [14]. The domain-specific knowledge incorporated with IML approaches refers to information not present in the input data. The three most prominent forms of knowledge incorporated using IML in medical applications [14], ranked by decreasing order of importance, are: 1) spatial invariances (e.g., [15]) widely used for image processing, 2) probabilistic relations (e.g., [16]) and 3) knowledge graphs (e.g., [17]). These representations can be integrated into ML models in various ways, such as by modifying the input data, altering the model structure, incorporating knowledge into the loss function, or comparing model outputs with known constraints [4]. Adding penalty terms to the loss function allows the model to balance between fitting the data and adhering to known domain-specific rules. This approach is particularly relevant in cases where learning purely from data may conflict with established domain knowledge. For instance, [18] introduced a Physics-Guided Neural Network (PGNN) incorporating physical rules between water temperature, depth, and density in DNN training ensuring consistency with mechanistic understanding. Similarly, [19] proposed a Domain-Adapted Neural Network (DANN) to integrate domain knowledge for monotonicity, valid ranges and approximation constraints. While DANNs are broadly applicable in fields like physics, engineering, and systems modeling—where knowledge can be expressed as structural constraints or approximate functions—they are less suited to epidemiology, where prior knowledge typically exists as statistical effect-size estimates from previous studies rather than explicit functional forms. Leveraging these effect sizes as a source of prior knowledge offers a natural way to guide learning in this context. Our approach, SEANN, is designed to do exactly this by incorporating PESs into the computation of epidemiological risk scores. In epidemiological studies, PESs are considered to be among the strongest levels of confidence for a factor’s relationship with health at a population level [20]. Despite their significance, there has been limited research on incorporating them into informed machine learning procedures, particularly using deep neural networks (DNNs). To our knowledge, our approach is the first to achieve this. The most similar work, introduced by Neri et al. in 2022 [9], proposed to integrate PESs into a naive Bayes model for the computation of health risk scores, the CArdiovascular LIterature-Based Risk Algorithm (CALIBRA). While their approach can combine input data with literature estimates to learn health relationships, it is limited to naive Bayes models and cannot be directly applied to other types of machine learning procedures. 3 Method This section introduces SEANN. We first define the general setting of integrating PESs as soft constraints via additional terms to the loss function and then detail the implementation of this approach for three types of PESs: standardized regression coefficients, odds ratios, and risk ratios. Given a set of p observed variables P, and a subset Q⊆P, with |Q|=q of these variables for which we have an effect size estimate value to use as enforced external knowledge, we define X, an n×p input matrix of n observations, and V, a vector of q effect size estimates values. Similarly to previous works (e.g., [21,22]), the general principle of our method, described in Eq 1, consists in adding a term to the loss function ℒ for each meta-heuristic to incorporate. Where ℒpred is the convex function used for the predictive task (e.g., mean squared error, cross-entropy, etc.) and θ the parameters of the model. ℒmeta is the convex loss function used to enforce the desired soft constraints for the neural network, i.e., to enforce the neural network to respect the PESs vector V. λ0 and λi are weights pondering the importance of each term, namely the predictive task and the ith constraint. They can be treated as hyperparameters and set to values that optimize the obtained predictive performances. However, for cases where learning plausible relationships, i.e., relationships aligned with known associations, is considered equally or more important than raw predictive power (e.g., imperfect input data, trustworthiness), a different approach should be used to settle the tradeoff between learning from the data (and optimizing performances) or learning associations observed in the literature. We propose choosing weight values proportional to the confidence in both the data and the external knowledge. Following a common principle in meta-analyze (e.g., [5]), we express this confidence using the sample size available in each study as well as the sample size available in the training data. The proposed weighting is calculated as follows: first, we define a confidence score ci associated with vi corresponding to the sample size of the ith meta-analysis. Similarly, we define a confidence score c0 for the input data X composed of n rows and p variables to be computed as n×p. Then, for c > 1, we estimate the final λ weights using a log scale relative normalization: This scheme ensures that terms associated with small confidence scores have a noticeable impact on the learning process compared to the others and are not entirely ignored. Depending on the type of PES considered, the loss function ℒmeta will be implemented differently. As effect estimates in the meta-analysis are typically represented either as OR, RR, or SRC, we express ℒmeta for those forms below. In all cases, the principle is to generate for each observation a slightly perturbated copy of it with an increment h—called the perturbation— for each variable in Q, to measure the difference between the expected change in the target value according to our PESs and the observed change in our model, and to penalize this difference. A numerical example is provided in Supporting information S1 Text. 3.1 Case of a standardized regression coefficient Let us first consider, for simplicity, a single SRC, called βi, that we want to integrate into the training of a DNN. This βi would be either directly extracted in a domain-specific literature study from a uni/multi-variate linear regression model or would summarize several similar effects in a meta-analysis. Considering a multivariate linear regression model defined as: With 𝐙∈ℝm1×m2 an input matrix and β, a vector of SRCs, βi∈β,1≤i≤m2. Then the expected change in the target values according to βi when modifying the corresponding input variable zi with a perturbation step h is: Where 𝐙zi+h denotes the matrix obtained from the input matrix Z by perturbing its i-th column, denoted as zi, through the addition of a quantity h, where h∈ℝ⧵{0}. To integrate βi within SEANN, we enforce a similar relationship between our model’s predicted values fθ(𝐗) and predicted values with perturbed inputs fθ(𝐗xi+h) as a soft constraint, i.e., fθ(𝐗)=fθ(𝐗xi+h)−βih, by penalizing the deviation from this equality. For a vector V of SRCs derived from the literature, with vi∈V, the ith element of V, we propose the following formulation for the training loss term ℒmeta in the context of SRC integration: Where 𝐗k denotes the kth row vector of matrix X. fθ is the output of the neural network with parameters θ, n the number of data points (i.e., batch size), and hi∈ℝ⧵{0} a perturbation parameter. In this case, as SRCs (similar to other PESs) are constant regardless of input data Z, we can theoretically use any value other than 0. For simplicity, we use hi = 1 for every SRCs to integrate. In a hypothetical, more general case of a constraint to integrate as a function of X, h would be taken as the smallest possible. 3.2 Case of an odds-ratio The approach we proposed in this section, while mathematically correct, can suffer from numerical instability during the training (cf., Eq 3). In paper 3, we addressed this issue by generalizing Eq 2 instead. Similar to Sect 3.1, let us consider a single OR, referred to as (ORi=eβi), that we want to integrate into the training process of a DNN. This OR would be extracted from logistic regression models in a meta-analysis. Considering a multivariate logistic regression model defined as: With 𝐙∈ℝm1×m2 an input matrix and β, a vector of m2 log-odds coefficients, βi∈β,1≤i≤m2. We can express the change in Logit(pβ) when modifying the input variable zi associated with βi with a perturbation step h. This difference is independent of other variables in Z. Thus, we can express the corresponding relationship between the predicted values on inputs Z with and without modifying the corresponding input variable zi with a perturbation step h: For a vector V of log-odds coefficients derived from the literature, with vi∈V, the ith element of V, we propose the following formulation for the training loss term ℒmeta in the context of OR integration: Where pθ is the probability given by the neural network with parameters θ, n the number of data points (i.e., the batch size) and h∈ℝ⧵{0} a perturbation parameter. Similar to Sect 3.1, as a given OR is constant for every corresponding z, h can theoretically take any values other than 0. However, within SEANN, we fix h to keep the quantities within the exponential terms small and enhance numerical stability during the learning process. We define: 3.3 Case of a risk ratio Following the same principle, let’s define the integration of a single PES encoded as a risk ratio. Considering a negative binomial regression model defined as: With 𝐙∈ℝm1×m2 an input matrix and β, a vector of log-estimates, βi∈β,1≤i≤m2. Then the expected change in the target values according to βi when modifying the corresponding input variable zi with a perturbation step h is: Where 𝐙zi+h denotes the matrix obtained from the input matrix Z by perturbing its i-th column, denoted as zi, through the addition of a quantity h, where h∈ℝ⧵{0}. To integrate βi within SEANN, we enforce a similar relationship between our model’s predicted values fθ(𝐗) and predicted values with perturbed inputs fθ(𝐗xi+h) as a soft constraint, i.e., fθ(𝐗)=fθ(𝐗xi+h)e−βih, by penalizing the deviation from this equality. For a vector V of log-estimates derived from the literature, with vi∈V, the ith element of V, we propose the following formulation for the training loss term ℒmeta in the context of RR integration: Where 𝐗k denotes the kth row vector of matrix X. fθ is the output of the neural network with parameters θ, n the number of data points (i.e., batch size), and hi∈ℝ⧵{0} a perturbation parameter. Similar to Sect 3.2, in order to keep the quantities within the exponential term small and enhance numerical stability, we recommend using: To demonstrate the potential of the approach, we rely on synthetic data that emulate different scenarios. In each experiment, we compare two DNNs, identical in all aspects but with the inclusion of our modified loss and external knowledge. For the sake of simplicity, we use and compare basic multilayer perceptrons. The approach can be directly usable with more complex feedforward neural architectures (e.g., convolutional networks, residual networks), and the benefits highlighted in this study should apply to other neural configurations. In the following, we call SEANN the model that implements our approach and the agnostic DNN the reference. 4 Experimental validation 4.1 Data scenario To illustrate the relevance in real applications, we introduce an intuitive fictional example composed of 1) a target variable y representing the risk of developing a disease or the strength of symptoms and 2) several variables contributing to the outcome y according to a dose-response relationship. To make the experiment more intuitive, we take a well-known confounding factor in environmental epidemiology: Reasonable fish intake tends to decrease health-related risks (e.g., cardiovascular disease and cognitive decline), while mercury intake increases them. However, fish intake is an important driver of mercury exposure. Thus, we define two correlated variables, mercury (x1) and fish intake (x2), having an opposite effect on the target variable. The correlation between x1 and x2 is designed to emulate a confounding effect [23]. We also define two additional variables, namely perceived stress (x3) and body mass index (i.e., BMI, x4) uncorrelated with variables x1 and x2 but affecting y. x3 is linear and positively correlated with the outcome, while x4 has a nonlinear effect. In this simple scenario, we perform different experiments in which we test the benefits of incorporating PESs into eligible variables (i.e., x1, x2 and x3). We are interested not only in the networks’ predictive performances, but also in their ability to capture and restitute the input-output relationships that we encoded in the data. The experiments focus on SRCs and ORs, but we could use RRs in a similar manner. 4.1.1 Standardized regression coefficients. For the case where PESs are encoded as SRCs, we generate an input matrix X by sampling m = 1000 values from a multivariate Gaussian with mean 0 and covariance matrix (10.8000.810000100001). Target vector Y is generated from the additive function described in Eq 5 with β0=1, β1=1, β2=−2, β3=5 and β4=10. 4.1.2 Odds-ratios. Similar to the linear case, we generate a data matrix X with three variables, namely x1, x2, and x3, corresponding to mercury, fish intake, and perceived stress, respectively, to predict an outcome (i.e., a risk to develop a disease). X was generated by sampling m = 1000 values from a multivariate Gaussian with mean 0 and covariance matrix (10.800.810001). Target vector Y was generated from the function described in Eq 6 with β0=0, β1=1, β2=−2, β3=5. 4.1.3 Experimental design. We use a fully connected neural network (NN) with a single hidden layer for both SEANN and the agnostic model. Both NNs were implemented using Pytorch and trained with a batch size of 64 and a maximum number of epochs of 1000. Parameter optimization was achieved using Adam [24]. We standardize and split data into training (n=600), validation (n=200), and test (n=200) datasets. To reduce overfitting, we use early stopping (with patience 10) on the validation set. 4.1.4 Evaluation. To evaluate the correctness of extracted relationships, we propose a score, called ΔShap, to compute the distance between two dose-response relationships represented with Shapley values [25]. ΔShap is defined by the mean absolute error (MAE) calculated across Shapley values for a given marginal relationship that must be computed using the same background dataset. We can use it to compare the distance between a neural network-extracted relationship and a relationship admitted in the literature or, in this work, as we use synthetic data, to compare the distance between a neural network-extracted relationship and the true predictor-outcome relationship. The smaller the distance, the more we would consider a relationship to be scientifically plausible or in line with the true effect. In this work, we use the generative functions (i.e., Eq 5 and Eq 6) to compute Shapley values representing the reference relationships for ΔShap. Shapley values are approximated using the SHAP library [26] and systematically computed on the test sets. To evaluate the performance of models, we use the coefficient of determination (R2) score for regression tasks and the receiver operating characteristic curve’s (ROC) area under the curve (AUC) for binary classification tasks. 4.2 Experiment 1 In this experiment, we illustrate that SEANN can leverage external expert information to mitigate the poor quality of the data. To simulate data imperfection, we gradually increase the level of noise on all variables (i.e., x1, x2, x3, x4) and check that while the performance of the agnostic NN deteriorates quickly, our informed NN can retain most of it. Information was degraded differently for linear coefficients and odds ratios to illustrate two common scenarios. In the linear case, we added Gaussian noise to all input variables, with a mean of 0 and an increasing standard deviation. For odds ratios, missing values were generated completely at random with increasing proportions and imputed using a simple mean imputation. External knowledge was integrated in addition to the training data for every eligible predictor (i.e., β1, , β3 for x1, , x3 respectively). Better performances were obtained with SEANN both for the predictive task and the explainability of constrained relationships measured with ΔShap. No significant gains were observed for the nonlinear unconstrained variable (BMI). The results are displayed in Fig 1 for the linear coefficients and summarized in Table 1 for the odds ratios. Performance comparison of SEANN and an agnostic DNN under increasing noise (Experiment 1). The x-axis shows the standard deviation of Gaussian noise added to all predictors. The y-axes report (top panel) predictive performance (R2) and (bottom panel) the alignment of extracted dose–response relationships with the ground truth, quantified by ΔShap (lower is better). Across noise levels, SEANN maintains higher predictive performance and more accurate relationships than the agnostic DNN. 10.1371/journal.pone.0338400.t001Table 1Comparison of Performances depending on the proportion of imputed missing values in training and validation sets (experiment 1).ROC AUC is a measure of predictive performances, whereas Sum Δ Shap summarizes the quality of captured relationships across all predictors. As noise level increases, SEANN maintains higher predictive performance and more accurate relationships than the agnostic DNN.Percent missingAgnostic DNNSEANNROC AUCSum Δ ShapROC AUCSum Δ Shap00.9990.0830.9980.137250.9300.2620.9750.176500.9150.3180.950.204 Comparison of Performances depending on the proportion of imputed missing values in training and validation sets (experiment 1). ROC AUC is a measure of predictive performances, whereas Sum Δ Shap summarizes the quality of captured relationships across all predictors. As noise level increases, SEANN maintains higher predictive performance and more accurate relationships than the agnostic DNN. The results indicate that when the PESs encode the correct relationships between the input variables and the target outcome, the predictive performances obtained while training on imperfect data are more stable with SEANN and the captured relationships are more scientifically credible. From a health research perspective, this means that when epidemiological data are noisy or incomplete—a common situation in cohort studies—SEANN can still recover relationships that are consistent with established knowledge. For example, while a standard NN tends to distort exposure effects as data quality deteriorates, SEANN preserves plausible dose–response patterns for mercury, fish intake, and stress by leveraging literature-informed constraints. In practice, this reduces the likelihood of drawing misleading conclusions about exposure–outcome relationships and improves the ability to retrieve meaningful effects under imperfect data conditions. 4.3 Experiment 2 In this experiment, we focus on the quality of the relationships captured when external knowledge is integrated for a single predictor. The objective is to show that variables that do not benefit directly from external knowledge may nevertheless be better captured, thanks to corrections brought to the other variables. Similarly to the previous experiment, Gaussian noise was added to a single variable (i.e., fish intake, x2) with mean 0 and standard deviation 0.75, and 1.5 for the SRCs and ORs respectively, in both training and validation sets. Fig 2 show results with PESs encoded as SRCs. The most significant gain was observed for fish intake, the variable with integrated external knowledge. ΔShap (see definition in Sect 4.1.4) measured for this variable was 1.11 with the agnostic DNN and 0.04 with SEANN. A significant gain was also observed for mercury, the variable correlated with the informed one (1.02 for agnostic DNN to 0.53 for SEANN). Finally, no significant performance gains were observed for the remaining variables, i.e., those uncorrelated with the informed one (0.99 to 1.04 for perceived stress, 1.74 to 1.64 for BMI, with agnostic DNN and SEANN, respectively). Comparison of extracted relationships (Shapley values) between the agnostic DNN and SEANN (experiment 2). Only the beta coefficient for fish intake was added as external knowledge. SEANN not only recovers the expected protective effect of fish intake (close alignment with external knowledge), but also improves the capture of the correlated harmful effect of mercury exposure. By contrast, the agnostic DNN produces distorted or attenuated associations under the same noisy conditions. The results show that not only SEANN better captures relationships for features with corresponding external knowledge but also for non-informed features that are correlated with those externally informed. From a health research perspective, these results illustrate that SEANN not only improves the interpretation of exposures with prior knowledge but also enhances the reliability of correlated predictors, helping disentangle complex mixtures where exposures co-occur. In practice, this means that known dose–response patterns (e.g., for fish intake) can guide the recovery of related effects (e.g., mercury), reducing bias from collinearity and enabling more trustworthy interpretation of intertwined environmental exposures. A similar scenario is observed for odds ratios, with ΔShap down from 0.087 with the agnostic DNN to 0.050 with SEANN for mercury. 4.4 Experiment 3 In a last experiment, we simulate a setting where a confounding variable is missing from the data. A confounding variable is a predictor impacting both the outcome to predict and other predictor(s) of interest. In numerous contexts, including health science, it is challenging to collect all relevant variables to predict an outcome (and study their effects), and we can expect to have unseen confounders. We train both NNs with a missing variable (fish intake) and compare both the predictive performance and the quality of the extracted relationships on the test set. With SEANN, we integrate external knowledge for mercury alone (i.e., the variable correlated with the missing predictor), and we duplicate the mercury variable in both training and validation datasets, the copy being an unconstrained variable. The objective of this duplication is that 1) The constrained version captures what is known by external knowledge, and 2) The unconstrained version captures what comes from the unseen confounder. We observe (Fig 3) that without the constraint, mercury is incorrectly captured, with a ΔShap of 1.32. In this case, interpreting the Shapley values directly could lead to the misleading conclusion that mercury has a protective effect. On the contrary, with SEANN, the constraint allows the capture of the correct relation (ΔShap=0.013). Additionally, SEANN was able to capture part of the association with the missing variable (fish intake) using duplicated input data of mercury (ΔShap: 0.43). Minor improvements were also observed for other variables. Results show that SEANN can be used to better disentangle individual effects while estimating the effect of unknown confounding factors. Comparison of extracted relationships (Shapley values) between the agnostic DNN and SEANN in the presence of a missing confounder (fish intake, Experiment 3). Without external knowledge, the agnostic DNN misinterprets mercury as protective, leading to misleading conclusions. SEANN, by constraining the mercury effect with PES information, correctly recovers its harmful association and partially captures the influence of the missing fish intake variable through the duplicated feature. For odds ratios, the results are similar. SEANN better captured the mercury predictor, with ΔShap=0.128 for the agnostic DNN and 0.048 for SEANN. SEANN was also able to capture part of the association with the missing variable (i.e., fish intake) using the duplicated input data (ΔShap= 0.056). 5 Discussion In this paper, we propose a method to integrate the wealth of knowledge available in the scientific literature encoded as pooled effect sizes (PESs). While these representations are simple estimates unable to express complex relationships, they are easily understandable, can be aggregated across multiple studies, and are widely used in epidemiology. By integrating these traditional statistical measures into the deep learning process, our approach offers a tool to capture complex nonlinear relationships from data while leveraging simpler but well-established knowledge. Our findings complement existing informed-ML approaches that encode domain knowledge as structural or mechanistic soft constraints (e.g., physics-guided networks or monotonic/range constraints implemented via loss regularization). In contrast, while prior work has injected effect-size estimates into simpler models such as naïve Bayes (e.g., CALIBRA), to our knowledge no method has integrated pooled effect sizes directly into neural network training, despite their prominence in epidemiological research. Our experimental protocol demonstrates that, compared with an agnostic DNNs with similar architecture, our approach offers two main benefits. First, improved generalization to unseen data: when PESs encode relevant information that is weak or absent in the observed data, SEANN’s predictive performance degrades more slowly under noise and missingness, yielding higher out-of-sample accuracy. Second, improved alignment of extracted relationships with external knowledge: SEANN preserves epidemiologically plausible dose–response patterns for both informed and uninformed variables and can be used to better disentangle individual input-output relationships in the presence of collinearity. For real-world epidemiological studies, these benefits imply more reliable risk predictions across different cohorts, populations or sites and more stable, policy-relevant, effect estimates despite measurement error, limited sample sizes, and partial confounding. Beyond the controlled synthetic experiments presented in this study, SEANN has already been applied to real-world epidemiological data. In a recent study on hypertension risk in European adults [27], we integrated SEANN into the construction of environmental risk scores using observational cohort data (from the GCAT project [28]) combining diverse environmental, sociodemographic, and clinical exposures. In this high-dimensional and partially confounded setting, SEANN improved the quality of captured relationships compared with agnostic neural networks, confirming its practical utility for epidemiological modeling. Building on our work on the early-life exposome in HELIX [29,30], future work will extend SEANN to life-course risk modeling and to larger multi-site cohorts, with the goal of enhancing environmental clinical risk scores. Supporting information (TEX) 8 Jul 2025 PONE-D-25-02109SEANN: A Domain-Informed Neural Network for Epidemiological InsightsPLOS ONE Dear Dr. Cazabet, Thank you for submitting your manuscript to PLOS ONE. After careful consideration, we feel that it has merit but does not fully meet PLOS ONE’s publication criteria as it currently stands. Therefore, we invite you to submit a revised version of the manuscript that addresses the points raised during the review process. Please submit your revised manuscript by Aug 22 2025 11:59PM. If you will need more time than this to complete your revisions, please reply to this message or contact the journal office at plosone@plos.org. When you're ready to submit your revision, log on to https://www.editorialmanager.com/pone/ and select the 'Submissions Needing Revision' folder to locate your manuscript file. Please include the following items when submitting your revised manuscript: A rebuttal letter that responds to each point raised by the academic editor and reviewer(s). You should upload this letter as a separate file labeled 'Response to Reviewers'.A marked-up copy of your manuscript that highlights changes made to the original version. You should upload this as a separate file labeled 'Revised Manuscript with Track Changes'.An unmarked version of your revised paper without tracked changes. You should upload this as a separate file labeled 'Manuscript'. A rebuttal letter that responds to each point raised by the academic editor and reviewer(s). You should upload this letter as a separate file labeled 'Response to Reviewers'. A rebuttal letter that responds to each point raised by the academic editor and reviewer(s). You should upload this letter as a separate file labeled 'Response to Reviewers'. A marked-up copy of your manuscript that highlights changes made to the original version. You should upload this as a separate file labeled 'Revised Manuscript with Track Changes'. A marked-up copy of your manuscript that highlights changes made to the original version. You should upload this as a separate file labeled 'Revised Manuscript with Track Changes'. An unmarked version of your revised paper without tracked changes. You should upload this as a separate file labeled 'Manuscript'. An unmarked version of your revised paper without tracked changes. You should upload this as a separate file labeled 'Manuscript'. If you would like to make changes to your financial disclosure, please include your updated statement in your cover letter. Guidelines for resubmitting your figure files are available below the reviewer comments at the end of this letter. If applicable, we recommend that you deposit your laboratory protocols in protocols.io to enhance the reproducibility of your results. Protocols.io assigns your protocol its own identifier (DOI) so that it can be cited independently in the future. For instructions see: https://journals.plos.org/plosone/s/submission-guidelines#loc-laboratory-protocols. Additionally, PLOS ONE offers an option for publishing peer-reviewed Lab Protocol articles, which describe protocols hosted on protocols.io. Read more information on sharing protocols at https://plos.org/protocols?utm_medium=editorial-email&utm_source=authorletters&utm_campaign=protocols. We look forward to receiving your revised manuscript. Kind regards, Vijay Kumar Academic Editor PLOS ONE Journal Requirements: When submitting your revision, we need you to address these additional requirements. 1. Please ensure that your manuscript meets PLOS ONE's style requirements, including those for file naming. The PLOS ONE style templates can be found at https://journals.plos.org/plosone/s/file?id=wjVg/PLOSOne_formatting_sample_main_body.pdf and https://journals.plos.org/plosone/s/file?id=ba62/PLOSOne_formatting_sample_title_authors_affiliations.pdf 2. Please note that PLOS ONE has specific guidelines on code sharing for submissions in which author-generated code underpins the findings in the manuscript. In these cases, we expect all author-generated code to be made available without restrictions upon publication of the work. Please review our guidelines at https://journals.plos.org/plosone/s/materials-and-software-sharing#loc-sharing-code and ensure that your code is shared in a way that follows best practice and facilitates reproducibility and reuse. 3. Thank you for stating in your Funding Statement: “The research is partially funded by the Meersens Company, financing the salary of the Ph.D student (first author) (https://meersens.com). The sponsor played no specific role in the research related to this manuscript.” Please provide an amended statement that declares *all* the funding or sources of support (whether external or internal to your organization) received during this study, as detailed online in our guide for authors at http://journals.plos.org/plosone/s/submit-now. Please also include the statement “There was no additional external funding received for this study.” in your updated Funding Statement. Please include your amended Funding Statement within your cover letter. We will change the online submission form on your behalf. 4. Thank you for stating the following financial disclosure: “The research is partially funded by the Meersens Company, financing the salary of the Ph.D student (first author) (https://meersens.com). The sponsor played no specific role in the research related to this manuscript.” Please state what role the funders took in the study. If the funders had no role, please state: "The funders had no role in study design, data collection and analysis, decision to publish, or preparation of the manuscript." If this statement is not correct you must amend it as needed. Please include this amended Role of Funder statement in your cover letter; we will change the online submission form on your behalf. 5. We note that your Data Availability Statement is currently as follows: All relevant data are within the manuscript and its Supporting Information files. Please confirm at this time whether or not your submission contains all raw data required to replicate the results of your study. Authors must share the “minimal data set” for their submission. PLOS defines the minimal data set to consist of the data required to replicate all study findings reported in the article, as well as related metadata and methods (https://journals.plos.org/plosone/s/data-availability#loc-minimal-data-set-definition). For example, authors should submit the following data: - The values behind the means, standard deviations and other measures reported; - The values used to build graphs; - The points extracted from images for analysis. Authors do not need to submit their entire data set if only a portion of the data was used in the reported study. If your submission does not contain these data, please either upload them as Supporting Information files or deposit them to a stable, public repository and provide us with the relevant URLs, DOIs, or accession numbers. For a list of recommended repositories, please see https://journals.plos.org/plosone/s/recommended-repositories. If there are ethical or legal restrictions on sharing a de-identified data set, please explain them in detail (e.g., data contain potentially sensitive information, data are owned by a third-party organization, etc.) and who has imposed them (e.g., an ethics committee). Please also provide contact information for a data access committee, ethics committee, or other institutional body to which data requests may be sent. If data are owned by a third party, please indicate how others may request data access. [Note: HTML markup is below. Please do not edit.] Reviewers' comments: Reviewer's Responses to Questions Comments to the Author 1. Is the manuscript technically sound, and do the data support the conclusions? The manuscript must describe a technically sound piece of scientific research with data that supports the conclusions. Experiments must have been conducted rigorously, with appropriate controls, replication, and sample sizes. The conclusions must be drawn appropriately based on the data presented. Reviewer #1: Yes Reviewer #2: Yes Reviewer #3: No ********** 2. Has the statistical analysis been performed appropriately and rigorously? Reviewer #1: Yes Reviewer #2: Yes Reviewer #3: No ********** 3. Have the authors made all data underlying the findings in their manuscript fully available? The PLOS Data policy requires authors to make all data underlying the findings described in their manuscript fully available without restriction, with rare exception (please refer to the Data Availability Statement in the manuscript PDF file). The data should be provided as part of the manuscript or its supporting information, or deposited to a public repository. For example, in addition to summary statistics, the data points behind means, medians and variance measures should be available. If there are restrictions on publicly sharing data—e.g. participant privacy or use of data from a third party—those must be specified. Reviewer #1: Yes Reviewer #2: No Reviewer #3: Yes ********** 4. Is the manuscript presented in an intelligible fashion and written in standard English? PLOS ONE does not copyedit accepted manuscripts, so the language in submitted articles must be clear, correct, and unambiguous. Any typographical or grammatical errors should be corrected at revision, so please note any specific errors here. Reviewer #1: Yes Reviewer #2: Yes Reviewer #3: Yes ********** 5. Review Comments to the Author Please use the space provided to explain your answers to the questions above. You may also include additional comments for the author, including concerns about dual publication, research ethics, or publication ethics. (Please upload your review as an attachment if it exceeds 20,000 characters) Reviewer #1: General Comments: The manuscript presents a novel and relevant approach by introducing SEANN, a domain-informed neural network model that leverages pooled effect sizes (PES) to enhance prediction and interpretation in epidemiological datasets. The work is timely and contributes to the growing intersection of machine learning and health sciences, especially in low-data settings. The manuscript is written in good English; however, it requires minor revisions before it can be accepted for publication. Specific Comments: 1. The manuscript, while clearly written, lacks consistent formatting throughout. Please ensure that headings, equations, and paragraphs follow the journal’s style guide. Several in-line equations are not properly spaced and disrupt the flow of reading. 2. Authors should check for the typographical errors such as in the Introduction section: “Odd Rations (ORs)” should be revised to “Odds Ratios (ORs).” 3. Authors should check for citing relevant citations in the text for e.g. On Page 5, the following line is missing a relevant citation: "For a vector V of SRCs derived from the literature, with vi ∈ V, the ith element of V, the training loss term Lmeta is defined as..." Cite appropriate literature to support the use of the custom loss formulation. 4. While the analysis appears comprehensive, the discussion lacks sufficient depth. It is recommended that the authors elaborate more on how their results compare with prior methods and explain the implications of improved generalizability in real-world epidemiological studies. Reviewer #2: 1.This manuscript introduces a novel approach, SEANN (Summary Effects Adapted Neural Network), which integrates pooled effect sizes (PESs) such as Odds ratios, Risk ratios, and standardized regression coefficients directly into deep neural network training. This is a unique and timely contribution to informed machine learning, particularly relevant to fields like epidemiology, where data is often limited or noisy. SEANN helps improve both predictive performance and interpretability by embedding prior scientific knowledge into model training. 2.The methodology is technically solid and clearly presented. The paper explains how PESs are embedded via custom loss functions and how weighting is derived based on confidence (sample size). The treatment of three common types of PESs is thorough. However, while the math is sound, the use of complex symbols and notations (Example: perturbation steps, arrow symbols) might challenge readers outside machine learning. Adding a simpler example with numeric inputs could help clarify the non-expert audiences. 3.The experiments are well-structured and simulate realistic epidemiological challenges noise, missing data, and confounding. The use of synthetic data is reasonable for isolating the effects of the method. SEANN consistently outperforms standard DNNs in terms of both prediction and interpretability, especially when data is imperfect. The use of the SHAP metric to assess alignment with true relationships is particularly valuable. However, the study would benefit from including comparisons with other informed machine learning models, like PGNN or DANN, or at least discussing their relative merits. 4.SEANN enhances the interpretability of model predictions, aligning well with real-world needs in health research. For example, it corrects misleading associations when variables are confounded or missing. While the figures and quantitative results are strong, some brief narrative explanations of what those corrections mean in practice, especially for health researchers, would make the work more engaging and accessible. 5.This is a strong and original piece of work. To improve it further, the authors should consider: Including a real-world dataset example. Reviewer #3: While the paper presents a creative idea — integrating domain knowledge in the form of pooled effect sizes into a neural network via a modified loss function — the manuscript ultimately lacks sufficient scientific rigor and practical relevance to warrant publication in PLOS ONE. My detailed concerns are as follows: 1. Lack of Real Data: The manuscript claims to improve epidemiological modeling, yet it contains no experiments on real-world epidemiological datasets. Synthetic toy examples (e.g., fish, mercury, stress, BMI) are insufficient to demonstrate the utility of the approach. Without validation on observational datasets from real health studies, the results cannot be trusted or generalized. 2. Overclaiming Practical Use: The term “epidemiological insights” in the title and abstract is misleading. No insights about disease risk, exposure, or population health are produced in this study. The method is demonstrated only in artificial settings, using pre-defined effects, which defeats the purpose of “insight” discovery. 3. Weak Baselines and Limited Evaluation: The authors compare SEANN only to an “agnostic DNN” — essentially a DNN without the PES-informed loss — and omit any comparison to: - Classical epidemiological methods (e.g., regularized logistic regression with prior constraints), - Knowledge-regularized models in causal inference, - Other informed machine learning models in epidemiology (e.g., those incorporating DAGs or monotonic constraints). The absence of meaningful baselines diminishes the credibility of the claimed improvements. 4. Unconvincing Justification of PES Integration: The proposed method modifies the loss function using hand-selected perturbation values (e.g., h=1 or h=1/vi), but these choices are not theoretically or empirically justified. There is also no sensitivity analysis regarding how these perturbations affect training stability or generalization. Furthermore, integrating PES values — which are often heterogeneous and context-dependent — as “hard-coded” truth in synthetic experiments does not reflect how they behave in actual epidemiological studies. 5. Limited Contribution to Machine Learning or Epidemiology: While the paper claims novelty in the integration of PESs into neural networks, the broader field of informed ML has explored similar loss regularization strategies with more robust theory and applications. This work does not meaningfully advance the methodological or epidemiological literature beyond existing approaches. 6. Figures and Tables Are Minimal and Not Informative Enough: The visualizations do not adequately explain how SEANN outperforms baselines. Key graphs (e.g., SHAP plots) are not contextualized with respect to meaningful health variables, and only small performance differences are observed in some metrics. ********** 6. PLOS authors have the option to publish the peer review history of their article (what does this mean?). If published, this will include your full peer review and any attached files. If you choose “no”, your identity will remain anonymous but your review may still be made public. Do you want your identity to be public for this peer review? For information about this choice, including consent withdrawal, please see our Privacy Policy. Reviewer #1: No Reviewer #2: No Reviewer #3: No ********** [NOTE: If reviewer comments were submitted as an attachment file, they will be attached to this email and accessible via the submission site. Please log into your account, locate the manuscript record, and check for the action link "View Attachments". If this link does not appear, there are no attachment files.] While revising your submission, please upload your figure files to the Preflight Analysis and Conversion Engine (PACE) digital diagnostic tool, https://pacev2.apexcovantage.com/. PACE helps ensure that figures meet PLOS requirements. To use PACE, you must first register as a user. Registration is free. Then, login and navigate to the UPLOAD tab, where you will find detailed instructions on how to use the tool. If you encounter any issues or have any questions when using PACE, please email PLOS at figures@plos.org. Please note that Supporting Information files do not need this step. 21 Oct 2025 We uploaded a PDF document with detailed answers to reviewers and editor. Submitted filename: SEANNanswerReviewers.pdf 23 Nov 2025 SEANN: A Domain-Informed Neural Network for Epidemiological Insights PONE-D-25-02109R1 Dear Dr. Cazabet, We’re pleased to inform you that your manuscript has been judged scientifically suitable for publication and will be formally accepted for publication once it meets all outstanding technical requirements. Within one week, you’ll receive an e-mail detailing the required amendments. When these have been addressed, you’ll receive a formal acceptance letter and your manuscript will be scheduled for publication. An invoice will be generated when your article is formally accepted. Please note, if your institution has a publishing partnership with PLOS and your article meets the relevant criteria, all or part of your publication costs will be covered. Please make sure your user information is up-to-date by logging into Editorial Manager at Editorial Manager® and clicking the ‘Update My Information' link at the top of the page. For questions related to billing, please contact billing support. If your institution or institutions have a press office, please notify them about your upcoming paper to help maximize its impact. If they’ll be preparing press materials, please inform our press team as soon as possible -- no later than 48 hours after receiving the formal acceptance. Your manuscript will remain under strict press embargo until 2 pm Eastern Time on the date of publication. For more information, please contact onepress@plos.org. Kind regards, Vijay Kumar Academic Editor PLOS ONE Additional Editor Comments (optional): As I have considered that the authors appropriately addressed all the comments from the reviewers, the manuscript can be accepted. Reviewers' comments: PONE-D-25-02109R1 PLOS ONE Dear Dr. Cazabet, I'm pleased to inform you that your manuscript has been deemed suitable for publication in PLOS ONE. Congratulations! Your manuscript is now being handed over to our production team. At this stage, our production department will prepare your paper for publication. This includes ensuring the following: * All references, tables, and figures are properly cited * All relevant supporting information is included in the manuscript submission, * There are no issues that prevent the paper from being properly typeset You will receive further instructions from the production team, including instructions on how to review your proof when it is ready. Please keep in mind that we are working through a large volume of accepted articles, so please give us a few days to review your paper and let you know the next and final steps. Lastly, if your institution or institutions have a press office, please let them know about your upcoming paper now to help maximize its impact. If they'll be preparing press materials, please inform our press team within the next 48 hours. Your manuscript will remain under strict press embargo until 2 pm Eastern Time on the date of publication. For more information, please contact onepress@plos.org. You will receive an invoice from PLOS for your publication fee after your manuscript has reached the completed accept phase. If you receive an email requesting payment before acceptance or for any other service, this may be a phishing scheme. Learn how to identify phishing emails and protect your accounts at https://explore.plos.org/phishing. If we can help with anything else, please email us at customercare@plos.org. Thank you for submitting your work to PLOS ONE and supporting open access. Kind regards, PLOS ONE Editorial Office Staff on behalf of Dr. Vijay Kumar Academic Editor PLOS ONE