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

arXiv:2610.11408 · cs.LG · Submitted 2026-10-08 · Read on arXiv

Listen

Transcript

Introduction to the show: ident: AI Radio. Generated commentary on the latest Artificial Intelligence papers.

Tom: Today's paper: "MC-TRCM: Observation-Aware Recursive Fusion for Incomplete Mobile and Wearable Mental-Health Feature Views".

Jane: The gist The Modality-Conditioned Temporal Recursive Context Model (MC-TRCM) preserves each feature source as a separate token and incorporates missingness as part of the input context to model incomplete mobile…

Tom: First, who's behind it and why it matters.

Title and authors: Tom: So, let's talk about who wrote this: Wentao Wang, Lifeng Han, Zining Ren, Hengyu Zhong, and Guangyu Zou from Dalian University of Technology and Hong Kong Polytechnic University. They put it under the title Observation-Aware Recursive Fusion for Incomplete Mobile and Wearable Mental-Health Feature Views.

Jane: That title tells you a lot about what they’re trying to achieve: making sure the model is aware of what data is actually there, which is crucial when dealing with incomplete mobile or wearable mental health information.

Lu: What's interesting about this work, as the paper says on page one, is that they frame these public datasets not as raw streams, but as feature-view learning problems where observed values and availability patterns are modeled jointly through source tokens and missingness projections > MC-TRCM: Observation-Aware Recursive Fusion for Incomplete Mobile and Wearable Mental-Health Feature Views #pg1.

Meng: So instead of just feeding the model a table of features, they're proposing this specific way to structure the learning problem so it understands how those different feature sources relate to each other and their missing parts.

Lalam: It’s about making sure the AI doesn't just look at what it sees, but also learns about what it *doesn't* see, which is really important for real-world data where things are always incomplete.

The paper's summary: Tom: When we look at the summary on page two of MC-TRCM, they explain that depression assessment still relies a lot on self-report and clinical encounters, and these gaps can leave clinically important intervals partially observed > MC-TRCM: Observation-Aware Recursive Fusion for Incomplete Mobile and Wearable Mental-Health Feature Views #pg2.

Jane: They point out that mobile sensing adds behavioral context between assessments, but that context isn't a direct measure of depression itself, so the absence of a proxy can actually be informative.

Lu: They are looking at datasets like StudentLife or DepreST-CAT, which release processed views mixing communication summaries and mobility statistics with symptom history and clinical variables > MC-TRCM: Observation-Aware Recursive Fusion for Incomplete Mobile and Wearable Mental-Health Feature Views #pg2.

Meng: The problem they're solving is that these releases mix so many different things—demographics, sleep, activity—and the missingness in any one of those pieces can be a key signal.

Lalam: It’s about modeling this whole collection of features as a feature view problem where the model has to jointly learn from what's observed and what's absent > MC-TRCM: Observation-Aware Recursive Fusion for Incomplete Mobile and Wearable Mental-Health Feature Views #pg1.

The paper's improvements: Tom: Now let’s look at the actual contributions on page three, because that’s where they detail how their observation-aware formulation and recursive architecture actually work > MC-TRCM: Observation-Aware Recursive Fusion for Incomplete Mobile and Wearable Mental-Health Feature Views #pg3.

Jane: They suggest using source tokens to preserve each feature source as a separate entity, which helps keep the data structure intact even when things are missing.

Lu: They introduce a mechanism where observed sources get encoded with values and missingness summaries, absent sources use learned absence tokens, and then they condition the fusion using dataset and task embeddings through something called FiLM > MC-TRCM: Observation-Aware Recursive Fusion for Incomplete Mobile and Wearable Mental-Health Feature Views #pg1.

Meng: And on top of that, they use a recursive decoding process where at each step the decoder combines the task-conditioned fused representation with the previous prediction feedback to update its state > MC-TRCM: Observation-Aware Recursive Fusion for Incomplete Mobile and Wearable Mental-Health Feature Views #pg10.

Lalam: This recursive refinement means they aren't just making one big guess; they are iteratively refining the answer over several steps, which helps stabilize the prediction process > MC-TRCM: Observation-Aware Recursive Fusion for Incomplete Mobile and Wearable Mental-Health Feature Views #pg10.

Conclusion: Tom: So to wrap up on this MC-TRCM paper, they’ve shown that by modeling missingness explicitly and using this recursive structure, they achieve the lowest mean absolute error on both DepreST-CAT severity endpoints among the models they tested > MC-TRCM: Observation-Aware Recursive Fusion for Incomplete Mobile and Wearable Mental-Health Feature Views #pg3.

Jane: It confirms that for continuous severity estimation, this observation-aware modeling approach gives a tangible reduction in error, specifically reducing MAE by zero point one eight one PHQ-nine points and zero point two one seven GAD-seven points compared to the strongest tabular references > MC-TRCM: Observation-Aware Recursive Fusion for Incomplete Mobile and Wearable Mental-Health Feature Views #pg10.

Lu: The ablation studies also show that both the context-conditioned fusion using FiLM and the explicit source absence tokens are important, meaning you need both parts to get the full benefit of this approach > MC-TRCM: Observation-Aware Recursive Fusion for Incomplete Mobile and Wearable Mental-Health Feature Views #pg10.

Meng: The paper does have a limitation they state: the classification results are task-dependent, with rounded ties or small numerical differences from the strongest references, which means performance varies depending on which specific endpoint you’re trying to predict > MC-TRCM: Observation-Aware Recursive Fusion for Incomplete Mobile and Wearable Mental-Health Feature Views #pg10.

Lalam: It’s a solid architecture for handling incomplete views, and it clearly separates the effects of the source modeling from the effects of just having released symptom or clinical context > MC-TRCM: Observation-Aware Recursive Fusion for Incomplete Mobile and Wearable Mental-Health Feature Views #pg10.

Tom: So, this MC-TRCM paper gives us a structured observation-aware architecture for handling those messy mobile and wearable mental health feature views, showing how to learn jointly from recorded values, source identity, and the pattern of available and missing sources > MC-TRCM: Observation-Aware Recursive Fusion for Incomplete Mobile and Wearable Mental-Health Feature Views #pg1.

Jane: It’s a deep dive into how we can make sense of these complex, incomplete data sets by treating every feature source as a separate token and incorporating missingness as part of the input context > MC-TRCM: Observation-Aware Recursive Fusion for Incomplete Mobile and Wearable Mental-Health Feature Views #pg1.

Lu: It shows that conditioning with dataset embeddings via FiLM can adapt the representation dynamically across different cohorts or tasks, which is a really flexible concept > MC-TRCM: Observation-Aware Recursive Fusion for Incomplete Mobile and Wearable Mental-Health Feature Views #pg10.

Meng: From an engineering standpoint, it’s about creating a robust framework where you can see exactly how much signal comes from passive sensing versus the actual clinical context > MC-TRCM: Observation-Aware Recursive Fusion for Incomplete Mobile and Wearable Mental-Health Feature Views #pg10.

Lalam: It really moves us forward in how we build AI that can handle the real mess of multimodal, real-time health data, showing how recursive refinement can improve the final output > MC-TRCM: Observation-Aware Recursive Fusion for Incomplete Mobile and Wearable Mental-Health Feature Views #pg10.

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

Dalian University of Technology · The Hong Kong Polytechnic University · Southwest University

cs.LG

Submitted: 2026-10-08

Updated: 2026-10-08

Comments: Accepted to ICONIP 2026. Code is available at https://github.com/Botwwt/MC-TRCM

Code: https://github.com/Botwwt/MC-TRCM

License: http://arxiv.org/licenses/nonexclusive-distrib/1.0/

Importance score: 86/100

The gist: The gist The Modality-Conditioned Temporal Recursive Context Model (MC-TRCM) preserves each feature source as a separate token and incorporates missingness as part of the input context to model

Key concepts

Modality Tokenization: Each feature source (like sensor values or demographics) is converted into a token. A gate mechanism then decides how much of the observed value and the missingness information should contribute to this token, creating a rich representation for every piece of data.
Missingness-Aware Gating: This component uses a learned gate ($G_m$) to control the influence of observed features ($ ilde{x}_{a,m}$) and missingness summaries ($q_{a,m}$). When data is missing, the model learns to use an 'absence token' instead of the actual value.
Recursive Decoding: The model generates its final output through multiple refinement steps. At each step, it combines task information and feedback from previous predictions to update its internal state. This iterative process allows for a more nuanced and accurate final prediction ($ ilde{y}_{a, au}$)

Summary

The gist The Modality-Conditioned Temporal Recursive Context Model (MC-TRCM) preserves each feature source as a separate token and incorporates missingness as part of the input context to model incomplete mobile and wearable mental-health feature views.

Problem Formulation

The core modeling challenge is to learn jointly from recorded values, source identity, and the pattern of available and missing sources in public datasets such as StudentLife, DepreST-CAT, Prediction of Severity Change-Depression (PSYCHE-D), Depresjon, and OBF-Psychiatric <ref:2610.11408#pg6>. The model observes standardized modality values x˜a,m, missingness summaries qa,m, and an observed indicator sa,m ∈ 0 or 1 <ref:2610.11408#pg4>. The learning problem is multitask prediction over regression, binary, ordinal, and multiclass endpoints while respecting endpoint-specific label availability <ref:2610.11408#pg6>.

Feature Views and Endpoints

The model organizes feature columns into five sources: sensor values, missingness, symptom context, static demographics, and clinical context <ref:2610.11408#pg6>. Sensor values include activity, sleep, communication, phone use, and mobility summaries when available <ref:2610.11408#pg6>. Missingness features include observed indicators, missing ratios, recency proxies, and modality masks <ref:2610.11408#pg6>. Symptom context includes temporally prior variables such as baseline PHQ-9 in PSYCHE-D or pre-study PHQ-9 in StudentLife <ref:2610.11408#pg6>. Static demographics include released age and sex variables <ref:2610.11408#pg6>. Clinical context includes prior treatment or baseline clinical variables when they are available before the modeled label <ref:2610.11408#pg6>. The primary benchmark contains six predefined endpoints: DepreST-CAT PHQ-9 severity, DepreST-CAT GAD-7 severity, DepreST-CAT PHQ-9 category, DepreST-CAT GAD-7 category, PSYCHE-D PHQ-change binary, and PSYCHED PHQ-change multiclass <ref:2610.11408#pg6>.

MC-TRCM Architecture

The MC-TRCM architecture involves several key components designed to handle feature heterogeneity and missingness <ref:2610.11408#pg6>.

** Modality tokenization and missingness-aware gating:**

The model encodes each feature source as a tokenized view, where an encoder Em maps standardized observed features into a shared token space, and a gate Gm controls how much of that observed-value representation should enter the token <ref:2610.11408#pg7>. The formula for the token is ta,m = (σ(Gm([˜xa,m; qa,m])) ⊙ Em(˜xa,m) + Amqa,m + em, sa,m = 1 <ref:2610.11408#pg8> or um + Amqa,m + em <ref:2610.11408#pg8>. When sa,m = 0 the token is formed from the learned absence token <ref:2610.11408#pg8>.

** FiLM dataset and task conditioning:**

Dataset and task embeddings form a context vector ca,τ = [e d d(a); e τ τ] <ref:2610.11408#pg9>. The reported configuration applies FiLM both before token fusion and after Transformer fusion, as selected on validation data and fixed before test evaluation <ref:2610.11408#pg10>.

** Recursive decoding:**

The decoder produces the task output over several refinement steps <ref:2610.11408#pg10>. At each step, it combines the task-conditioned fused representation, the task identity, and feedback from the previous prediction to update an internal decoder state <ref:2610.11408#pg10>. The final step gives the reported prediction yˆa,τ = o(Kd)a,τ <ref:2610.11408#pg10>.

Training Objective and Optimization

The training objective involves accumulating losses over recursive steps using normalized increasing weights wk = k/PKd <ref:2610.11408#pg10>. The loss is defined as Ltask = (sum exp(-λτ) X Kd k=1 wkL(k)τ + λτ) / TBX <ref:2610.11408#pg10>. Regression tasks compare Huber, MSE, and hybrid MAEHuber losses in validation search <ref:2610.11408#pg10>. Binary and multiclass tasks include classbalanced cross entropy, focal loss, and label-smoothing options when selected by validation <ref:2610.11408#pg10>. The shared encoder is trained across related but nonidentical clinical endpoints, producing gradient geometry in which PHQ-9 severity, GAD-7 severity, category prediction, and PHQ-change classification may emphasize different symptom dimensions <ref:2610.11408#pg10>.

Main Benchmark Results

MC-TRCM is most clearly beneficial for continuous severity estimation <ref:2610.11408#pg10>. It reduces MAE by 0.181 PHQ-9 points and 0.217 GAD-7 points relative to the strongest tabular references, indicating that observation-aware source modeling is particularly useful when the target is symptom magnitude rather than a discrete category <ref:2610.11408#pg10>. Classification results are more endpoint-dependent <ref:2610.11408#pg10>. MC-TRCM matches the best rounded GAD-7 category BA and is numerically highest on PSYCHE-D multiclass by 0.002 BA, but remains close to the strongest references on PHQ9 category and PSYCHE-D binary prediction <ref:2610.11408#pg10>.

Ablation Studies

Component ablations show that both context-conditioned fusion and explicit source absence are important for incomplete feature-view learning <ref:2610.11408#pg10>. The largest degradations come from removing FiLM and absence tokens, which increases severity MAE by 0.136 and 0.125, respectively, and also lowers category BA <ref:2610.11408#pg10>. Missingness projection and recursive refinement provide smaller but consistent gains <ref:2610.11408#pg10>.

Conclusion

MC-TRCM provides a structured observation-aware architecture for heterogeneous mobile and wearable mental-health feature views <ref:2610.11408#pg10>. Under validation-only selection, it achieves the lowest error on DepreST-CAT PHQ-9 and GAD-7 severity estimation and shows endpoint-specific classification behavior relative to strong tabular and neural references <ref:2610.11408#pg10>. Component ablations support the contributions of FiLM conditioning, missingness modeling, and recursive refinement, while feature-source accounting separates architecture effects from released symptom, static, and clinical context <ref:2610.11408#pg10>.

How it works

** Modality tokenization and missingness-aware gating:**

The model encodes each feature source as a tokenized view, where an encoder Em maps standardized observed features into a shared token space, and a gate Gm controls how much of that observed-value representation should enter the token <ref:2610.11408#pg8>. The formula for the token is ta,m = (σ(Gm([˜xa,m; qa,m])) ⊙ Em(˜xa,m) + Amqa,m + em, sa,m = 1 <ref:2610.11408#pg8> or um + Amqa,m + em <ref:2610.11408#pg8>.