PaperlayerView exact revision on arXiv

Source-linked AI summary

MC-TRCM: Observation-Aware Recursive Fusion for Incomplete Mobile and Wearable Mental-Health Feature Views

Wentao Wang, Lifeng Han, Zining Ren, Hengyu Zhong, Guangyu Zou

arXiv:2610.11408v1cs.LG

TL;DR

Public mobile and wearable mental-health releases often provide processed feature views rather than synchronized raw streams, combining indirect sensor proxies with clinical and symptom context. MC-TRCM preserves feature sources and availability patterns while conditioning fusion on dataset and task. It achieves the lowest MAE on DepreST-CAT PHQ-9 and GAD-7 severity, with endpoint-specific classification behavior.

  • Problem

    Public mental-health feature releases combine indirect sensor proxies with symptoms, clinical context, and availability indicators, so modeling must account for recorded values, source identity, and missingness.

    Source: §1 · p. 2

  • Method

    MC-TRCM represents feature sources as separate tokens, models absence and missingness, conditions fusion on dataset and task, and recursively refines endpoint predictions.

    Source: §1 · pp. 2–3

  • Results

    MC-TRCM reduced MAE by 0.181 PHQ-9 points and 0.217 GAD-7 points versus the strongest tabular references, with endpoint-specific classification behavior.

    Source: §1 · p. 3, §6 · p. 13

  • Takeaways & Limitations

    Feature-source controls show that symptom and clinical context account for substantial signal in prepared releases, clarifying what passive sensing and missingness views add.

    Source: §5 · p. 13, §6 · p. 13

    The current source-factorized tokenization uses preaggregated anchor summaries; extension to unaligned raw sensor time series depends on those streams being available.

    Source: §5 · p. 13

Abstract

from arXiv · show

Public mobile and wearable mental-health datasets often provide summarized feature tables rather than synchronized raw sensor streams. In these releases, each anchor corresponds to a survey or label time and may combine phone or wearable summaries, prior symptom scores, demographics, clinical variables, and source-availability indicators. We propose the Modality-Conditioned Temporal Recursive Context Model (MC-TRCM), which preserves each feature source as a separate token and incorporates missingness as part of the input context. Observed sources are encoded with values and missingness summaries, absent sources use learned absence tokens, dataset and task embeddings condition fusion, and a recursive prediction head refines each output over validation-selected steps. We evaluated MC-TRCM on six predefined endpoints from DepreST-CAT and Prediction of Severity Change-Depression (PSYCHE-D) using participant-level splits and validation-only model selection. MC-TRCM achieved the lowest mean absolute error on DepreST-CAT Patient Health Questionnaire-9 (PHQ-9) and Generalized Anxiety Disorder-7 (GAD-7) severity, improving over the best tabular reference by 0.181 and 0.217 scale points. Classification endpoints showed task-dependent behavior: MC-TRCM matched the best rounded GAD-7 category balanced accuracy, was numerically highest by 0.002 balanced-accuracy points on PSYCHE-D multiclass prediction, and remained close to the strongest references on PHQ-9 category and PSYCHE-D binary prediction. Ablations support Feature-wise Linear Modulation, absence tokens, missingness projections, and recursive refinement, while calibration and feature-source controls characterize endpoint behavior. Our code is available at https://github.com/Botwwt/MC-TRCM.

1 Introduction

Public mental-health releases combine indirect sensor proxies with symptom, clinical, and availability features, creating a feature-view learning problem. MC-TRCM preserves source identity and missingness, and its strongest reported gains are on severity estimation.

1 Introduction

  • Public releases mix processed sensor views with symptom history, demographics, clinical context, and availability indicators, making source identity and missingness part of the modeling challenge.Missing fields may reflect non-wear, disengagement, hospitalization, privacy settings, or release constraints; sensing features are indirect behavioral proxies.Source: §1 · p. 2
  • MC-TRCM represents each feature source as a token and uses missingness-aware gates, learned absence tokens, and projections, with dataset- and task-conditioned FiLM and recursive prediction refinement.The model is designed to use measured values and availability patterns without collapsing sources into one flat vector.Source: §1 · p. 2
  • The study formulates incomplete public releases as feature-view learning, jointly modeling observed values, source identities, and availability patterns through source tokens and recursive decoding.Its reported contributions include a leakage-controlled benchmark with explicit feature-source accounting.Source: §1 · p. 3
  • The evaluation covers six DepreST-CAT and PSYCHE-D endpoints with participant-level splits and validation-only model selection, calibration, thresholding, and recursive-depth choices.It includes matched feature-source controls to distinguish sensor and missingness views from symptom, static, and clinical context.Source: §1 · pp. 2–3
  • MC-TRCM reduced MAE by 0.181 PHQ-9 points and 0.217 GAD-7 points versus the strongest tabular references, while classification results were task-dependent.The study reports the lowest MAE on both DepreST-CAT severity endpoints.Source: §1 · p. 3

2 Related Work

Prior work spans public digital-phenotyping benchmarks, tabular and neural predictors, and methods for missingness and conditioned fusion. The section motivates endpoint-aware metrics and preserving source and availability structure in processed feature views.

2.1 Digital Phenotyping Benchmarks, Metrics, and Calibration

  • Mobile and wearable studies use activity, sleep, mobility, phone-use, and communication signals, while cohort and protocol differences motivate reproducible benchmarks and endpoint-appropriate metrics.Severity is evaluated in scale points, while imbalanced category and change labels call for class-balanced metrics; calibration also matters.Source: §2.1 · p. 3

2.2 Generalization, Missingness, and Feature Views

  • Cross-study generalization can shift across cohorts, devices, years, and sensing protocols, while matrix imputation can obscure the distinction between low, partially reliable, and unavailable sources.MC-TRCM keeps sources separate, represents availability explicitly, and evaluates matched feature-source controls.Source: §2.2 · p. 4

2.3 Tabular, Neural, and Conditioned Fusion Models

  • Tabular models and neural sequence or set encoders are established references, but processed mental-health releases call for preserving modality, endpoint, and availability structure even without raw streams.MC-TRCM combines source tokens and modality absence with dataset- and task-conditioned fusion and recursive task-state updates; FiLM conditions representations, while PCGrad addresses task gradient conflict.Source: §2.3 · p. 4

3 Methods

MC-TRCM frames participant-level prediction from incomplete, heterogeneous feature views as source-token learning with explicit availability context. Its dataset- and task-conditioned fusion, recursive decoding, and validation-controlled training and evaluation are designed for six predefined severity and classification endpoints.

3.1 Problem Formulation

  • Repeated anchors from each participant stay in a single split, and each endpoint is predicted only when its label is available.An anchor may be a survey time, label time, or released processed row paired with features available at or before it.Source: §3.1 · p. 4

3.2 Feature Views and Endpoints

  • The six primary endpoints combine DepreST-CAT severity and category tasks with PSYCHE-D PHQ-change binary and multiclass prediction.Feature views cover sensor values, missingness, prior symptom context, static demographics, and clinical context.Source: §3.2 · p. 5

3.3 MC-TRCM Architecture

  • Each feature source becomes a token, with observed values gated alongside missingness summaries and absent sources represented by learned absence tokens.For partially missing observed sources, values are training-only imputed and standardized while masks, missing ratios, and recency proxies enter through missingness summaries.Source: §3.3 · p. 6
  • Dataset and task embeddings condition fusion, while a task-specific recursive decoder feeds prior predictions forward and reports the output at a validation-selected depth.The recursive state is an internal prediction-refinement variable, not a clinically validated latent state.Source: §3.3 · pp. 7–8

3.4 Training Objective and Optimization

  • Training supervises successive recursive predictions with increasing weights on later refinements and validation-selects task losses and conflict handling.The loss choices include Huber, MSE, hybrid MAE-Huber, class-balanced cross entropy, focal loss, label smoothing, or ordinal decoding, as selected by validation.Source: §3.4 · p. 8

3.5 Baselines, Calibration, and Metrics

  • The benchmark compares tabular and neural references under shared splits and preprocessing, reporting MAE for severity, balanced accuracy for classification, and calibration measures.Thresholding, class-bias correction, and temperature scaling are fitted on validation data, excluding test labels.Source: §3.5 · p. 9

3.6 Ablations and Robustness

  • Five-seed ablations remove individual components, while recursive-depth sensitivity, calibration comparisons, and matched feature-source controls assess robustness.The ablated components include FiLM, absence tokens, missingness projections, PCGrad, task conditioning, and recursive refinement.Source: §3.6 · p. 10

4 Experiments and Results

Across six predefined endpoints, MC-TRCM’s clearest primary-benchmark advantage is in severity estimation, while classification results vary by endpoint. Feature-source controls, calibration and recursive-depth selection, and component ablations further characterize its performance.

4.1 Main Benchmark Results

  • Classification is endpoint-dependent: MC-TRCM matches the best rounded GAD-7 category BA, is numerically highest on PSYCHE-D multiclass by 0.002 BA, yet remains close to strongest references on PHQ-9 category and PSYCHE-D binary prediction.Source: §4.1 · p. 10
  • 0.181 PHQ-9 points and 0.217 GAD-7 points lower MAE than the strongest tabular references are MC-TRCM’s clearest primary-benchmark gains.Table 1 reports pre-specified comparator families after validation-selected hyperparameters and recursive depth.Source: §4.1 · p. 10, Table 1 · p. 11
  • Auxiliary results include StudentLife PHQ-9 severity MAE of 3.812 ± 0.302 versus 4.916 ± 0.229, while OBF-Psychiatric reaches BA = 1.000 for both MC-TRCM and LightGBM.Table 2 covers auxiliary endpoints outside the primary benchmark because held-out sets are small or saturated.Source: §4.1 · p. 11, Table 2 · p. 11

4.2 Calibration and Recursive Depth

  • SSC-only results are competitive with the full model and slightly exceed it for DepreST-CAT PHQ-9 category; sensor and missingness views alone are weaker on most endpoints.The feature-source figure also indicates endpoint-dependent contributions from sensor and missingness views.Source: §4.2 · pp. 11–12, Figure 3 · p. 12
  • For DepreST-CAT PHQ-9 category prediction, MC-TRCM reduces Brier score and ECE by 0.098 and 0.108 relative to LightGBM.The figure reports endpoint-specific calibration behavior.Source: §4.2 · p. 12, Figure 3 · p. 12
  • Validation selects recursive depth Kd = 8 for DepreST-CAT and Kd = 4 for PSYCHE-D, where performance then declines.The selection scores are within-dataset validation scores, not test metrics or cross-dataset comparisons.Source: §4.2 · p. 12, Table 3 · p. 12

4.3 Ablation Studies

  • Removing FiLM or absence tokens increases severity MAE by 0.136 and 0.125, respectively; missingness projection and recursive refinement provide smaller but consistent gains.Five-seed ablations also show a minor effect from PCGrad and mixed average task-conditioning effects on categorical endpoints.Source: §4.3 · pp. 12–13, Table 4 · p. 13

5 Limitations and Future Work

Calibration and feature-source controls reveal endpoint-specific evaluation behavior, while future work targets uncertainty, source regularization, cross-cohort adaptation, and raw sensor streams.

5 Limitations and Future Work

  • Calibration varies by endpoint: MC-TRCM improves DepreST-CAT PHQ-9 category calibration, while LightGBM and EBM lead on the other listed classification endpoints.LightGBM has lower calibration error on GAD-7 category prediction; EBM has lower Brier score and ECE on both PSYCHE-D classification endpoints.Source: §5 · p. 13
  • Feature-source controls show that symptom and clinical context account for substantial signal in prepared public releases, so performance is not attributable to sensors alone.The controls clarify what passive sensing and missingness views add.Source: §5 · p. 13
  • Future work includes uncertainty estimation, feature-source regularization, and domain adaptation across cohorts and release protocols.The source-factorized tokenization could also extend from preaggregated anchor summaries to unaligned raw sensor time series when available.Source: §5 · p. 13

6 Conclusion

MC-TRCM offers a structured observation-aware architecture for heterogeneous mobile and wearable feature views, with its clearest results in severity estimation and support from component ablations.

6 Conclusion

  • MC-TRCM achieves the lowest error on DepreST-CAT PHQ-9 and GAD-7 severity estimation, with endpoint-specific classification behavior against tabular and neural references.Feature-source accounting separates architecture effects from released symptom, static, and clinical context.Source: §6 · p. 13
  • MC-TRCM preserves heterogeneous mobile and wearable feature views through a structured, observation-aware architecture.The conclusion identifies dataset and task conditioning, missingness modeling, and recursive refinement as components supported by ablations.Source: §6 · p. 13
  • Component ablations support FiLM conditioning, missingness modeling, and recursive refinement, while feature-source accounting distinguishes architecture effects from released context.Source: §6 · p. 13
Loading 2610.11408v1…