PaperlayerView exact revision on arXiv

Source-linked AI summary

Representation Is Not Enough: Body-Localized Thermal Evidence for Contactless Stress and Craving Sensing in Opioid Use Disorder

Sachin Deb, Harshit Sharma, Asif Salekin

arXiv:2608.16087v1cs.CVcs.LG

TL;DR

Contactless stress sensing must localize the body regions and moments supporting coarse clip-level labels without wearable supervision. FABLE-Therm preserves localized evidence through weakly supervised, multi-encoder fusion, achieving 0.938 AUROC while transferring to craving and exposing participant-level deployment gaps.

  • Problem

    Contactless thermal stress sensing must identify which body regions and moments support a coarse clip-level label after wearable supervision is removed.

  • Method

    FABLE-Therm treats body-region trajectories from frozen encoders as one weakly supervised multiple-instance problem, preserving spatial and temporal evidence until final embedding-level fusion.

  • Results

    0.938 AUROC on unseen participants, 0.752 AUROC for self-reported craving, and participant-resolved analysis showed representation improvements recover only about half the deployment gap.

  • Takeaways & Limitations

    Preserving localized evidence supports accurate contactless stress sensing, craving-related signal recovery, and analysis of whether deployment failures reflect representation or person-to-person heterogeneity.

  • Takeaways & Limitations

    Small samples yield wide bootstrap intervals, while craving and StressNet evaluations are feasibility checks rather than benchmarks.

Abstract

from arXiv · show

Removing wearables from physiological monitoring also removes their supervision: the signal indicating where and when a stress response occurred. Contactless stress sensing therefore becomes a weakly supervised evidence-localization problem, where a clip-level label must be traced to the body regions and moments that produced it. We address this with FABLE-Therm, a weakly supervised architecture that preserves localized evidence across body regions, time, and encoder-specific representations until the final decision. FABLE-Therm fuses frozen foundation-model encoders at the embedding level, with theory explaining why localized fusion can outperform feature concatenation and prediction averaging. We study this problem in opioid use disorder (OUD), where stress is a major relapse trigger and sustained wearable use can be difficult during early recovery. Using fixed thermal video, FABLE-Therm achieves 0.938 AUROC on held-out participants, and its learned representation transfers to self-reported craving, providing, to our knowledge, the first evidence that craving can be recovered from contactless thermal video. Localized evidence also enables participant-level analysis of deployment failure. We find that improving representation alone is insufficient for equitable deployment: additional data from the underserved group would recover only about half of the cohort gap, while the remainder reflects person-to-person heterogeneity. This modality-agnostic decomposition applies to models with identifiable subpopulations. Together with the first cohort-structured contactless thermal OUD benchmark, our results show that preserving localized evidence supports both accurate sensing and principled analysis of who a model fails and why.

1 Introduction

Contactless thermal stress sensing is framed as weakly supervised evidence localization because removing wearables removes supervision about where and when responses occur. FABLE-Therm preserves localized evidence across regions and moments until final fusion, achieving strong participant-held-out stress performance and transferring to craving sensing.

  • Motivation and problem: Removing wearable biomarkers leaves one coarse clip label, requiring models to identify which body regions and moments support a stress response.This reframes contactless sensing as weakly supervised evidence localization.
  • Core method: FABLE-Therm preserves each region’s and moment’s potential evidence under weak supervision, deciding which evidence matters only at the final decision.Its design integrates frozen foundation-model encoders into a single multiple-instance learning framework while retaining spatial and temporal localization.
  • Stress sensing: 0.938 AUROC on unseen participants, 0.203 above the strongest of 27 baselines, demonstrates FABLE-Therm’s stress-sensing performance.The result is reported for evaluation on held-out participants.
  • Craving transfer: 0.752 AUROC on self-reported craving outperforms the strongest craving baseline, demonstrating that contactless thermal sensing carries craving-related signals.The model was retrained on craving labels, which represent a related but distinct relapse-associated construct.
  • Problem formulation and architecture: The paper formulates thermal stress or craving sensing as identifying which regions, over which moments, support a thermal-video-window label.FABLE-Therm uses E frozen encoders and keeps evidence tied to (view, region) instances until a learned gate fuses embeddings.

2 Related Work

Prior work spans contact-based stress sensing, camera-based physiological monitoring, and thermal psychophysiology, while highlighting supervision, subject dependence, and labeling challenges. Related foundations include frozen self-supervised encoders and multiple-instance learning, although existing weak-supervision methods assume static instances from one encoder.

  • Stress sensing and thermal psychophysiology: Contact-based physiological sensors established stress recognition, while camera-based methods reduce contact but alter the supervision signal.Camera-based approaches use rPPG or video paired with contact references.
  • Stress sensing and thermal psychophysiology: Subject dependence and inconsistent labeling remain recurring barriers in camera-based stress sensing.These barriers are identified across camera-based physiological monitoring literature.
  • Stress sensing and thermal psychophysiology: In opioid use disorder, stress reactivity and craving are related but distinct: stress often induces craving, yet craving can also occur without stress and varies across individuals.The passage distinguishes stress-induced craving from craving that occurs independently of stress.
  • Foundation models, MIL, and Disparity in health: Frozen self-supervised encoders transfer across domains, thermal foundation models have emerged, and multiple-instance learning is standard for weak supervision.Existing multiple-instance learning methods assume static instances from a single encoder, an assumption later relaxed in Section 4.

3 Problem Formulation

The problem is weakly supervised: each thermal video window has one clip-level label, so the model must recover where and when the response occurs and test whether that evidence generalizes across cohorts. The formulation explicitly models latent body region, response time, and contextual nuisance variation, while evaluating performance under cohort shift rather than relying only on aggregate AUROC.

  • Problem formulation: Each participant contributes a thermal video window with one binary elicitation-condition label, requiring prediction from weak clip-level supervision.The label is assigned to the window rather than to individual frames or body regions.
  • Problem formulation: The clip-level label does not identify where or when the response occurs, so the model must recover latent evidence whose generalization is tested with person-disjoint splits.Person-disjoint evaluation tests whether localized evidence transfers across cohorts.
  • Learning under weak supervision: The latent structure includes the responding body region, response-development time, and surrounding body context needed to separate physiological responses from nuisance variation.For example, uniform warming may be ambient, whereas a warm site against a stable body may indicate a response.
  • Learning under weak supervision: Global pooling and fixed ROIs can fail silently when physiology differs across cohorts, motivating explicit modeling of the latent structure.The formulation avoids removing localization ambiguity by assumption.
  • Evaluation under cohort shift: Aggregate accuracy alone can hide underserved-cohort failures, so each cohort is evaluated both when included in training and when held out entirely.The corpus is majority-Control, with 42 Control participants versus 29 OUD participants, allowing average performance to improve while OUD performance degrades.

4 FABLE-Therm Architecture

FABLE-Therm preserves localized thermal evidence across time, body regions, and frozen encoder representations before making a window-level decision. Its shared, embedding-level aggregator combines temporal dynamics, spatial context, gated multiple-instance pooling, and participant-centered calibration.

  • Localized tiling: The model masks and tiles each thermal window into N = 1 + RC tokens, retaining one global token and an R × C grid of local regions.The segmenter suppresses background regions that may correlate with room or cohort.
  • Frozen encoders: With 67 participants, three frozen encoders produce regional trajectories, while only the aggregator is trained to limit identity memorization.The encoders are AnyThermal, InfMAE, and ImageNet-ViT, selected for complementary pretraining.
  • Branch aggregation: Each encoder branch preserves when evidence appears, order-sensitive trajectory changes, body-relative context, and weak-label credit assignment across regions before pooling.Temporal descriptors include mean, standard deviation, signed slope, and mean successive change; spatial attention precedes gated MIL pooling.
  • Embedding-level fusion: A per-window gate fuses branch embeddings as a single product-bag attention-MIL operation, with region weights selecting evidence and encoder weights selecting which view to trust.The routing cue is an uncalibrated softplus scalar, not a confidence or uncertainty estimate.
  • Fusion rationale: Sharing the aggregator makes cross-encoder embedding addition well-posed, while embedding-level fusion preserves structured evidence that feature- and posterior-level fusion can lose.Untying the operator causes a large accuracy drop, according to the cited ablation result.
  • Training and calibration: The aggregator minimizes six loss terms, and participant centering uses up to 24 unlabeled windows for approximately 1.5-min transductive calibration rather than cold-start detection.Centering makes classification respond to personal deviations instead of absolute baseline temperature.

5 Evaluation Design

The evaluation separates population, dataset, and endpoint generalization across distinct datasets and person-disjoint protocols. It evaluates transfer to unseen participants, external StressNet data, and craving self-report while controlling cohort composition and participant-level uncertainty.

  • Evaluation scope: Distinct datasets test population, dataset, and construct generalization rather than relying on one result for all claims.
  • Research questions: RQ1 tests transfer to unseen participants and Control-to-OUD application; RQ2 tests external transfer to StressNet; RQ3 tests transfer from stress labels to self-reported craving.
  • Datasets: 67 participants across 71 sessions comprised Dataset I: 42 control adults and 25 individuals with OUD, including 29 OUD sessions.Four OUD participants were recorded twice, producing 29 OUD sessions; analyses use session-level counts.
  • Datasets: 16,273 windows comprised Dataset I, with 11,410 Control and 4,863 OUD windows; stress labels were 77.5%, and cohort priors were 76.7% Control versus 79.2% OUD.The stress prior was protocol-driven, so label composition could not explain the cohort gap.
  • Protocols: Four person-disjoint settings use fixed splits shared by every method, including Test-Both, Test-Control, Test-OUD, and Control→OUD.Test-Both contains 3,306 windows; Control→OUD trains and validates on Control only and tests on all 29 OUD sessions.
  • Uncertainty: Participant-level cluster bootstrap with B=2000 resamples participants rather than windows for transfer claims to unseen people.Paired comparisons reuse participant draws and report percentile 95% CIs and one-sided empirical bootstrap tail probabilities; seed variation is reported separately as seed SD.

6 Results

FABLE-Therm achieves strong held-out-participant stress sensing and transfers to craving prediction from contactless thermal video. Its localized, shared embedding fusion also exposes a cohort gap split between representation and person-to-person heterogeneity, limiting data-only equity remedies.

  • Stress sensing: 0.938 AUROC on held-out participants exceeds the strongest general baseline, TMC, at 0.735, while Control→OUD performance remains 0.771.Random Forest reaches 0.802 accuracy but only 0.612 AUROC, and LoRA never exceeds 0.686.
  • Cohort-gap decomposition: 0.158 representation gain and 0.169 heterogeneity gap split the cohort deficit roughly 48/52, showing that within-cohort variability is comparably important to missing OUD data.On the same six held-out OUD participants, adding OUD training data raises AUROC from 0.652 to 0.810, versus a 0.979 Control ceiling.
  • Fusion ablation: Embedding-level fusion with FABLE-Therm reaches 0.938 AUROC, outperforming feature-level Concat-3 at 0.736 and posterior-level TMC at 0.735 on Test-Both.The controlled comparison holds frozen inputs, splits, and seeds fixed while varying only the aggregator.
  • Fusion ablation: 0.744 ±.012 AUROC for the untied competitor versus 0.938 ±.004 for FABLE-Therm on Test-Both shows that shared operators remain necessary despite 1.80× the trainable parameters.The same comparison gives 0.674 ±.053 versus 0.810 ±.002 on Test-OUD, with substantially greater seed variability for the untied model.
  • Cross-dataset transfer: 0.808 ± 0.112 AUROC across three subject-disjoint StressNet folds indicates operation across a second sensor, stressor, and grid resolution, with substantial participant-dependent variation.The reported value is an average across folds rather than a selected best fold.
  • Craving transfer: 0.752 AUROC on self-reported craving exceeds MIMMO, providing initial evidence that OUD craving is recoverable from contactless thermal video.The feasibility probe uses 29 OUD sessions from 25 individuals, containing 4,863 windows but only 172 independent labels.

7 Threats to Validity

The paper bounds its claims by noting that the headline comparison is system-level and that small protocol samples produce wide bootstrap intervals. Ablation isolates the aggregator and shared-operator choice, but not every architectural component.

  • Ablation scope: The headline comparison is system-level because FABLE-Therm uses an Optuna-tuned recipe while most baselines use a vanilla loop.Table G.1 directly isolates the aggregator and shared-operator choice; the statistics path, spatial attention, and auxiliary losses are argued rather than individually ablated.
  • Small samples: 6–29 sessions per protocol yield wide bootstrap intervals, so uncertainty is reported explicitly.

8 Responsible Deployment and Release

The paper frames FABLE-Therm as opt-in clinical decision support rather than diagnosis or a substitute for clinical judgment, with deployment requiring participant-specific calibration and cohort-specific thresholds. Release is governed by explicit non-use restrictions, reproducible artifacts, and protected access to raw clinical recordings.

  • Intended deployment: The intended role is opt-in clinical decision support—a revocable check-in—not diagnosis, substance-use inference, or replacement for clinical judgment.Thermal video is body-derived and person-linkable rather than anonymous.
  • Intended deployment: Deployment requires short participant-specific calibration because the method is transductive and not a cold-start detector.It also requires cohort-specific operating thresholds rather than a universal 0.5.
  • Prohibited uses: Surveillance, coercive decisions, and non-consensual inference are prohibited across law enforcement, probation or parole, employment, insurance, and eligibility decisions.These restrictions are enforced through the artifact license rather than left as recommendations.
  • Release and governance: The release includes code, frozen representations, scripts, fixed splits, per-seed predictions, attribution checkpoints, and a machine-readable non-use license, while raw clinical recordings remain IRB-governed.The released artifacts support reproducing results without raw video and enable the external-transfer experiment end to end because StressNet is public.

9 Conclusion … C.1 Dataset I (our collection).

The paper presents contactless thermal stress and craving sensing as a localized, weakly supervised evidence problem, while showing that deployment gaps reflect both representation limits and participant heterogeneity. Its appendices specify the method, formal scope, reproducibility resources, and cohort-structured Dataset I.

  • 9 Conclusion: FABLE-Therm preserves body- and time-localized evidence, making contactless predictions more accurate and interrogable while exposing participant-resolved deployment failures.On the cohort, improving representation alone recovers only about half of the deployment gap.
  • Appendix: Code and test data are available for reproduction, while the full dataset is planned for release upon paper acceptance.The appendix organizes supporting material by the first use of each appendix reference.
  • A Extended Related Work: The related-work appendix omits headline scores because differing stressors, targets, populations, splits, and label definitions make direct ranking invalid.The conceptual framing also draws on thermal psychophysiology, addiction medicine, health equity, construct validity, and mental-privacy concerns.
  • B.1 Window construction: Each analysis window contains 25 target times spanning 4.8 seconds, and timestamp matching rejects mismatched, repeated, or out-of-interval frames before branch fusion.This alignment ensures encoder branches describe the same physical instants.
  • B.2 Thermal rendering and its physical scope: Per-region min–max rendering enables pretrained image encoders but removes calibrated absolute temperature, uniform regional warming, and cross-region temperature comparisons.Frozen encoders retain time and region axes, then encoder-specific projections align their otherwise incompatible coordinate systems.
  • B.3 Projection; B.5 Spatial contextualization and MIL pooling; B.6 Fusion policy. b⋆= P: Temporal descriptors expose level, dispersion, signed trend, and short-term variation before learned mixing, while spatial MIL localizes evidence after contextualizing regions.The learned gate recomputes encoder view weights for every window from branch embeddings and quality cues.
  • B.8 Participant centering; B.9 Formal statements and their scope: Participant centering requires up to 24 unlabeled calibration windows for held-out participants, and formal results limit claims about order sensitivity, product-bag allocation, gauge symmetry, and posterior fusion.The formal scope explicitly distinguishes content-only attention from order-sensitive alternatives and identifies posterior fusion as a rank-E bottleneck.
  • C Datasets and Splits: Dataset I uses 16,273 windows from 71 sessions and 67 participants, with 77.5% labeled stress and separate Control and OUD cohorts.The corpus contains 42 Control participants and 25 unique OUD individuals across 29 sessions.

C.2 Dataset Characteristics. · C.3 Dataset II (StressNet). · C.4 Dataset III (craving).

The datasets combine heterogeneous session-level craving labels with contactless physiological windows from StressNet and OUD craving recordings. Their construction preserves temporal context and requires subject-independent evaluation because session labels are often single-class and craving unfolds gradually.

  • C.2 Dataset Characteristics.: Craving rates vary across the seven retained task codes, from approximately 0.40 for Count to 0.68 for Bad-memory recall, while no task is purely one class.Task identity therefore does not determine the craving label.
  • C.2 Dataset Characteristics.: 16 of 29 sessions (55%) are single-class, with 10 containing no craving windows and 6 entirely craving, despite comparable session window totals of 45–59.Only 13 sessions contain both craving labels, so pooled near-balance obscures strong session-level heterogeneity.
  • C.2 Dataset Characteristics.: Because labels are heterogeneous across sessions and tasks, supervision is sparse and subject-independent splitting is essential to avoid session-level label leakage.A random-window split could place windows sharing a single-class session label into both training and test sets.
  • C.3 Dataset II (StressNet).: 305 trials from 34 participants form StressNet under a cold-pressor protocol, with session 2 labeled no stress and session 3 labeled stress.Each example spans 100 source frames over 13.3 s, sampled to 25 frames, and uses 97 regions represented at 768 dimensions.
  • C.2 Dataset Characteristics.: 55.5% of 4,863 pooled windows are craving, but session-level prevalence remains highly heterogeneous across the 29 OUD sessions.The corpus contains 2,698 craving and 2,165 non-craving windows.
  • C.4 Dataset III (craving).: 4,863 OUD windows from 29 sessions and 25 individuals are reused for craving prediction by replacing the stress target with per-block binary craving self-reports.The dataset contains 2,698 craving and 2,165 non-craving windows.
  • C.4 Dataset III (craving).: Craving-related thermal signatures unfold over tens of seconds, so longer windows capture more autonomic dynamics and align the input extent with task-block supervision.The craving label is assigned once per task block rather than at a single instant.

D Full Baseline Results (Primary Corpus) … E.2 Why the original StressNet number is not the right comparison target.

Across person-disjoint primary-corpus evaluations, AUROC is the controlled comparison, with FABLE-Therm reaching 0.938 and outperforming general baselines while accuracy can be prior-driven. The appendices also disclose protocol, implementation, and comparison-target caveats, including why the original StressNet headline is not directly comparable.

  • D Full Baseline Results (Primary Corpus): AUROC is the controlled comparison because FABLE-Therm uses a validation-selected threshold while most baselines retain 0.5.Accuracy and F1 therefore describe saved operating points rather than a calibration comparison.
  • D.1 Reading Table D.2.: 0.802 and 0.793 baseline accuracy coexist with 0.612 and 0.621 AUROC for Random Forest and ExtraTrees, showing that accuracy is prior-driven here.These methods post the highest baseline accuracy but near-chance ranking.
  • D.1 Reading Table D.2.: TMC is the best Test-Both baseline, whereas the median rule is best for Control-to-OUD transfer, so within-distribution ranking does not predict cross-cohort ranking.The protocols can invert method ordering.
  • D.1 Reading Table D.2.: 0.938 AUROC is FABLE-Therm’s primary-corpus result, with a 0.203 margin over the best general baseline’s 0.735 AUROC.The margin is arithmetically 0.938 − 0.735.
  • D.2 LoRA harness disclosure.: LoRA uses rank-8, α=16, dropout-0.05 adapters and a trainable Linear(19,200, 2) head, with approximately 270k–295k trainable parameters against 86.6M–96.7M frozen backbones.Training uses AdamW with lr 10−3, weight decay 10−4, class-weighted cross-entropy, bf16 precision, and early stopping.
  • D.2 LoRA harness disclosure.: Results within the fixed single-draw fold are controlled comparisons rather than stable dataset-wide rankings because fold ranking and participant identity are entangled.The split files remain identical and person-disjoint, but the LoRA family does not use the same training harness as elsewhere.
  • D.2 LoRA harness disclosure.: The full results report Test-Both and Control-to-OUD transfer as mean±std over 3 seeds, with exact positive-class accuracy floors of 0.788 and 0.792.The 0.775 value is the full-corpus rate, not either test-set floor.
  • E.2 Why the original StressNet number is not the right comparison target.: The original StressNet result used an 80%/10%/10% flat random split of individual subject-session trial files without subject grouping, unlike this person-disjoint evaluation.StressNet’s original supervision reconstructed an ECG-derived heat-emission/ISTI signal unavailable in this corpus, so it was evaluated directly on the task labels instead.

F Craving (RQ3) Full Results · F.1 Scope of the feasibility claim. · G Full Aggregator Comparison

FABLE-Therm provides the first reported decoding of self-reported craving from thermal video, achieving the strongest accuracy and balanced accuracy while showing that rankings differ by metric. The results support out-of-person craving associations at a usable operating point, and the accompanying tables define cohort-wise evaluation protocols and aggregator comparisons.

  • F.1 Scope of the feasibility claim.: AUROC is threshold-free, whereas accuracy, balanced accuracy, and F1 use a common fixed 0.5 threshold for the craving endpoint.LoRA and MIL-ablation heads are omitted for this endpoint, and positive prevalence is 55.5%.
  • F.1 Scope of the feasibility claim.: FABLE-Therm is reported as the first demonstration that self-reported craving can be decoded from thermal video.The claim is framed as feasibility evidence for this setting.
  • F.1 Scope of the feasibility claim.: 0.667 accuracy and 0.660 balanced accuracy are FABLE-Therm’s best values in the craving table.MIMMO is the closest baseline for balanced accuracy at 0.603.
  • F.1 Scope of the feasibility claim.: Table D.3 evaluates separately on held-out Test-Control and Test-OUD cohorts after training on both cohorts.Test-Control contains 8 subjects and 2,306 windows; Test-OUD contains 6 subjects and 1,000 windows.
  • F.1 Scope of the feasibility claim.: 0.970, 0.967, and 0.796 are AUROC values for AnyThermal, Concat-3, and the Min rule, exceeding FABLE-Therm’s 0.752 AUROC.The passage emphasizes that AUROC rankings and thresholded operating-point decisions measure different properties.
  • F.1 Scope of the feasibility claim.: The archived experiment establishes out-of-person ranking and operating-point associations with block-level craving reports.The passage presents this as the first such evidence for thermal video in this setting.

G.1 What these tables establish. … I.4 Population-level findings and their status.

Across matched evaluations, FABLE-Therm benefits from regional trajectories and late fusion, while its implementation preserves localized, participant-balanced evidence for attribution. Population-level maps show non-uniform regional contributions but also underscore anatomical heterogeneity across individuals.

  • G.1 What these tables establish.: Regional trajectories outperform static-vector instances across all tested MIL heads, and the shared three-encoder model is strongest on both cohorts.The full model reaches 0.938 AUROC on Test-Both and 0.810 on Test-OUD; the proposed single-encoder branch reaches 0.747 on Test-Both.
  • H.1 System: The system processes 25 sampled frames with three frozen encoders, updating only the aggregation network on a single NVIDIA RTX 5090.The encoders are AnyThermal, InfMAE, and a third timm-based feature artifact; hyperparameters are selected with Optuna 4.8.
  • H.2 Training: Training combines mixed precision, gradient clipping, cosine warmup, decoupled AdamW decay, selective stochastic weight averaging, validation threshold tuning, and validation-based early stopping.Gradient clipping uses max-norm 5.0, patience is 25, and evaluation is deterministic under seeds {42, 100, 2023}.
  • H.3 Search protocol: Optuna searches train-to-validation configurations over five trials of 200 epochs, after which the best trial is retrained for three seeds.The search space favors slower learning and stronger regularization because an unconstrained frozen-feature model overfits within one epoch.
  • H.3 Search protocol: The preliminary four-encoder comparison cannot isolate LanguageBind’s contribution because encoder count, gate mode, and hyperparameter tuning change simultaneously.The comparison uses the same person-disjoint split and three seeds, but the four-encoder model has uniform gating and untuned hyperparameters.
  • I.1 Visual pipeline and reading conventions: The attribution pipeline masks the background, partitions each frame into a fixed 6 × 8 grid plus a whole-frame token, and tracks each location across 25 frames.Attention weights indicate pooling emphasis, whereas signed grouped-SHAP values indicate whether a region raises or lowers the stress score.
  • I.2 Grouped Kernel-SHAP: A grouped Kernel-SHAP player jointly reveals or masks one token’s 25-frame, 768-channel trajectory across all three encoders, producing 49 interpretable players.The final fused logit is explained using zero in normalized feature space as the missing-feature reference.
  • I.3 Sampling: Attribution sampling draws two windows from every available participant–class stratum and averages within participants before aggregating across participants.The fixed person-disjoint test split contributes 54 windows from all 14 held-out participants, preventing participants with more windows from dominating.

I.5 Five held-out OUD case studies. … J.3 Tier 3 — raw recordings, governed request.

The case studies show that stress evidence is participant-specific and can reflect shortcut-like spatial patterns, while per-encoder analyses reveal complementary regional explanations. The reproducibility artifact provides public code and governed access paths for derived features and raw participant recordings.

  • I.5 Five held-out OUD case studies.: Each of the five OUD participants had a different strongest stress-supporting grid cell, with no cell repeated.The strongest cells were R4C7, R5C7, R6C7, R3C6, and R4C4 for O01, O02, O03, O05, and O06, respectively.
  • I.5 Five held-out OUD case studies.: Stress evidence differed by participant, spanning lateral trunk and forearm, right-margin motion-heavy, upper-body, and facial regions.This pattern is presented as a per-participant analogue of aggregate cohort differences in where attribution magnitude concentrates.
  • I.5 Five held-out OUD case studies.: For C02, peripheral negative effects drove a misclassification: R5C8 contributed −1.627 and R6C8 contributed −0.732, outweighing R6C7 at +0.690.The correctly classified window had p(stress) = 0.935, while the misclassified window had p(stress) = 0.159.
  • I.5 Five held-out OUD case studies.: The O04 comparison limits that shortcut interpretation: R5C4 contributed −0.942 near the lower torso rather than at an obviously empty boundary.The correctly classified window had p(stress) = 0.948, whereas the misclassified window had p(stress) = 0.097.
  • I.7 Per-encoder attribution and 25-frame window probabilities.: The three encoders produced similar whole-window stress probabilities but assigned evidence to different regional trajectories.Across the two examined windows, branch probabilities were 0.89–0.93; branch logit magnitudes are not calibrated across encoders, so attribution bars should be compared within encoder.
  • J.1 Tier 1 — fully public, no request required.: The artifact supports reproducibility at increasing data-access levels, with public code, fixed folds, evaluation scripts, and table-generation scripts available without contacting the authors.StressNet is public, and Tier 1 is described as the recommended entry point for applying the method to a new thermal corpus.
  • J.3 Tier 3 — raw recordings, governed request.: Derived features require consent and governance, while raw thermal recordings are human-participant data accessed through a documented request procedure.Tier 2 release requires IRB/consent authority, threat assessment, data-use terms, and removal of unnecessary metadata; raw recordings cannot be presumed public.

J.4 Applying the method to a new domain. … K.5 On “privacy-preserving”.

The method is portable across domains through region masks, tiling, per-region frozen encoders, and window labels, while its deployment requires strict governance, explicit non-uses, and caution about heterogeneous errors and privacy. The paper frames contactless thermal sensing as an advisory research prototype rather than a diagnostic, surveillance, or privacy-preserving system.

  • J.4 Applying the method to a new domain.: FABLE-Therm transfers to new domains using a person/object mask, configurable tiling, any frozen per-region encoder, and window labels.The token count is architecture-dependent; configurations with N=97 and N=49 demonstrate that the grid is not a hard-coded body model.
  • J.5 Operating-point provenance.: Table J.1 documents one validation-selected threshold applied identically to held-out Control and OUD cohorts, with participant-cluster bootstrap intervals and descriptive, noncausal comparisons.The threshold is 0.374906, and the table reports cohort-specific confusion counts while population, site, and cohort composition remain entangled.
  • K.1 Intended benefit pathway.: An opt-in, non-contact signal may enable low-burden check-ins when sensor wear, charging, or maintenance is impractical, but it should not diagnose stress or substance use.Its intended role is to prompt conversation or offer support without replacing self-report or clinical judgment.
  • K.2 Prohibited uses, as licence terms.: Law enforcement, supervision, employment monitoring, insurance, eligibility decisions, and covert or non-consensual monitoring are prohibited as licence conditions.The passage emphasizes that contactless sensing can operate without subject cooperation, making advisory language insufficient.
  • K.3 Conditions for a responsible pilot.: Responsible piloting requires revocable consent, participant benefit, limited retention or on-device processing, access and deletion controls, human review, validation, subgroup audits, and co-design.FABLE-Therm is explicitly identified as a research prototype that does not currently satisfy these requirements.
  • K.4 Failure modes and who bears them.: At the seed-42 operating point, OUD participants have higher observed false-positive and false-negative rates than Control participants, with wide intervals and no deployment validation.False positives may cause intervention or stigma, while false negatives may create false reassurance; per-person calibration is required for deployment.
  • K.5 On “privacy-preserving”.: Thermal imaging records a person-linkable, body-derived signal; subject-adversarial training discourages identity encoding but provides no formal privacy guarantee.The paper therefore makes no formal privacy claim and notes potential special-category and cognitive-biometric protections.
Loading 2608.16087v1…