When Adaptation Hurts: Split Sensitivity and Person-Level Negative Transfer in Federated Wearable Onboarding

• Unlabeled onboarding may worsen a predictor for people with small or class-incomplete calibration sets, while window-level evaluation can conceal correlated observations and fall false alarms.The introduction identifies recording leakage, dependent device endpoints, and repeated seeds as sources of understated uncertainty; fall sensitivity alone does not measure operational false alarms.Source: §1 · p. 2
• The onboarding validity contract combines federated source provenance, immutable participant splits, recording-disjoint calibration and evaluation, outer-person inference, matched zero-shot branches, and explicit calibration availability.Source: §1 · p. 2
• The frozen two-layer benchmark retains a 21-method broad screen and completes nine-user HHAR and eight-subject PAMAP2 rotations for six core strategies.Source: §1 · p. 2
• HHAR’s original frozen-person balanced accuracy exceeded every corresponding complete-rotation mean by 13.8–17.8 pp; paired uncertainty leaves displayed leader changes unresolved near-ties, not established superiority.Source: §1 · p. 2
• The failure analysis separates seed-averaged participant effects from descriptive seed–person realizations and examines finite-sample harm bounds, tail quality, availability, cost, and fall specificity.Source: §1 · p. 2
• The study is a retrospective, audited development benchmark—not an independent confirmatory superiority trial—and promotes no universal winner or safety claim.Earlier candidate and selector criteria were not met, and unfavorable results remain visible.Source: §1 · p. 2

2.1 Federated heterogeneity and personalization

• FedAvg aggregates local updates and FedProx modifies local objectives; personalized methods retain local layers [8], local normalization, meta-learn initialization, pair global and personal models, or learn strategies.Most personalization assumes target-client participation in training or labeled local data; this study instead fixes source training and considers adaptation from a short unlabeled stream for an unseen client.Source: §2.1 · p. 3

2.2 Test-time adaptation and federated onboarding

• TENT, EATA, and ATP motivate label-free adaptation; prior TTA work also identifies model-selection sensitivity, small-batch negative transfer, and unstable online updates.The novelty is not that negative transfer exists, but its person-level frequency and severity with a federatively trained source and unavailable evaluation recordings for calibration; BoTTA examines mobile and edge constraints.Source: §2.2 · p. 3

2.3 Federated wearable recognition and evaluation integrity

• The evaluation gap is federated wearable onboarding assessed with recording-disjoint calibration, outer-person inference where feasible, branch-matched harm, tail quality, calibration availability, fall false-positive windows, and artifact provenance.Wearable TTA is not new: OFTTA adapts online on the test stream, whereas this protocol uses a bounded calibration pool from separate recordings and a federated rather than pooled source.Source: §2.3 · p. 3

3 Problem Formulation and Statistical Contract

• Each target unit supplies unlabeled calibration data to adapt a source-trained checkpoint; evaluation uses disjoint recordings, with b = 32 as the frozen primary budget.Other calibration budgets are supporting sensitivity analyses.Source: §3 · p. 3
• For HHAR, balanced accuracy is averaged across device endpoints within each seed and then across seeds; dataset summaries and intervals are computed across people, not windows, endpoints, or seeds.Other datasets have one endpoint per outer subject or user.Source: §3 · p. 4
• The stress audit has 114 seed–person values per method from 38 people, which are repeated observations—not independent participants—and expose joint source-training and calibration variability without separating its sources.These stress-audit observations receive no population-level confidence interval.Source: §3 · p. 4
• Negative transfer is gain below zero, while material harm is gain below −2 pp; both are within-branch measures distinct from cross-method absolute balanced accuracy.Rates and worst observed gain are reported; an adaptation method can improve its own weak zero-shot state yet remain below a stronger source model.Source: §3 · p. 4
• Uncertainty uses deterministic 10,000-draw bootstrap resampling of outer units and paired participant-level differences; with only 6–9 units per headline domain, intervals are descriptive, not safety guarantees.When no material-harm events are observed, the contract reports the exact one-sided 95% Clopper–Pearson upper bound 1−0.051/n rather than treating zero observations as zero population risk.Source: §3 · p. 4

3.1 Onboarding validity contract

• The validity contract excludes target people from source optimization and normalization, separates physical calibration and evaluation recordings, bars evaluation-label access, and preserves unavailable budgets as missing.It also requires identifying each method’s source checkpoint and adaptation dependencies by fold and seed, and averaging endpoints and seeds before inference across held-out people.Source: §3.1 · p. 4
• Absolute post-onboarding balanced accuracy measures final predictor usefulness, while person-level gain measures improvement over that branch’s own zero-shot state; neither substitutes for the other.A largest mean gain does not by itself make a method best, and a positive mean gain does not establish safety.Source: §3.1 · p. 4

4.1 Datasets, preprocessing, and outer splits

• Frozen artifacts preserve sample, user, domain, and recording identities, with source-only normalization and overlap checks required before runs are reportable.Calibration uses 40% of eligible recordings; requested windows come only from calibration, and unavailable budgets are not filled from evaluation. UCI HAR was excluded for lacking auditable trial identities, while UP-Fall is a five-subject derivative with one descriptive outer unit.Source: §4.1 · pp. 4–5, Table 1 · p. 5

4.2 Source model and federated training

• A common temporal convolutional backbone and a fixed federated-training recipe define the source models, with FedBN substituting BatchNorm for GroupNorm.The TCN uses width 96, three residual depthwise–pointwise blocks, and a 128-dimensional embedding; source FL runs at most 200 rounds, and target evaluation labels never select checkpoints.Source: §4.2 · p. 5

4.3 Development selection and hyperparameter freeze

• The final five-domain effect estimates are development estimates, and the frozen 270-job HHAR/PAMAP2 extension is not an independent confirmation study.Earlier development outcomes informed the architecture, hyperparameters, primary budget, and paper framing; no post-rotation parameter was retuned.Source: §4.3 · p. 5

4.4 Core onboarding methods

• Six fully rotated core strategies comprise three static controls and three calibration-consuming mechanisms, including FedBN recalibration, an ATP-style entropy step, and Feature-only moment alignment.ATP-style updates approximately 56–60k target parameters; Feature-only freezes the network and classifier and uses no target-trainable parameters. FedBN recalibrates target BatchNorm statistics without gradients.Source: §4.4 · p. 6

4.5 Evidence layers and reporting outcomes

• The 837-job evidence union combines a 567-job broad layer with a predeclared 270-job rotation extension, and job count is execution coverage rather than statistical sample size.Reported measures include accuracy, macro F1, balanced accuracy, p10 and worst outer-unit BA, harm and negative-transfer rates, timing, and fall-specific classification measures.Source: §4.5 · p. 6

5.1 Coverage and evidence integrity

• All 837 declared jobs completed, and all 147 primary-budget method–domain rows include every eligible outer unit and all three seeds.The audit verified all 18 frozen derived files and passed with claim constraints; completed HHAR and PAMAP2 rotations support split-sensitivity analysis, while the other three headline domains corroborate utility and harm without newly expanded rotations.Source: §5.1 · pp. 6–7

5.2 RQ1: a single held-out person is not a stable benchmark

• 13.8–17.8 pp lower HHAR balanced accuracy across the nine-user means than the original-fold results shows how strongly the single-person split changes the benchmark.The displayed leader changes from parameter-matched FedPer to FedBN on HHAR and from FedPer-GH to parameter-matched FedPer on PAMAP2; paired leader-runner intervals include zero, and the nominal HHAR leader wins fewer than half of pairings.Source: §5.2 · p. 7, Table 2 · p. 7

5.3.1 Absolute post-onboarding quality

• Against FedAvg, Feature-only exceeds by +1.46 pp on HHAR (95% outer bootstrap CI [+0.45,+2.71] pp) and +2.36 pp on WISDM ([+0.05,+6.01] pp), but is −0.36 pp on SisFall.Source: §5.3.1 · p. 7

5.3.2 Within-branch gain and harm

• 10/114 Feature-only, 4/114 FedBN, and 22/114 ATP-style seed-person realizations are harmful, affecting eight, two, and 16 people, respectively.These are repeated seed-person realizations, not additional participants; the worst single realization is materially lower than the seed-averaged worst for every adaptive method. Mean-gain signs remain stable under leave-one-person-out for most domain-method combinations, but a stable positive mean is not a no-harm result.Source: §5.3.2 · pp. 8–9
• No adaptive core method achieves positive mean gain in every dataset and zero observed person-level losses greater than 2 pp.ATP-style improves HHAR, UniMiB, and WISDM (including +7.18 pp on WISDM) but loses −2.64 pp on PAMAP2; FedBN’s positive mean gains do not prevent one materially harmed PAMAP2 subject, while Feature-only has no seed-averaged decline beyond 2 pp but has a 7.6% exact one-sided 95% upper bound.Source: §5.3.2 · pp. 8–9

5.4 RQ3: mean, tail, and calibration quantity answer different questions

• The criterion changes the displayed leader on PAMAP2, SisFall, and UniMiB, while calibration curves can plateau or regress and higher-budget availability thins outside HHAR and PAMAP2.On UniMiB, parameter-matched FedPer has the largest mean while FedBN has the strongest observed lower tail; with only 6–9 outer units, p10 is an interpolated descriptive statistic, and budget is not an automatic safety gate.Source: §5.4 · p. 9, Figure 2 · p. 10

5.5 RQ4: fall sensitivity can conceal false-positive windows

• 92.9% sensitivity but 36.0% specificity for Feature-only on the single UP-Fall outer unit illustrates how fall sensitivity can conceal false-positive windows.On SisFall, parameter-matched FedPer has the largest displayed sensitivity and F1 while retaining 94.4% specificity; window counts do not establish false alarms per hour or event-level detection.Source: §5.5 · p. 9, Table 4 · p. 10

5.6 Measured cost and negative method-selection results

• At 32 calibration windows, ATP-style adapts 56.1–60.0k target parameters in 6.9–10.4 ms, while Feature-only trains no parameter and takes 45.2–50.1 ms for moment estimation and alignment.No domain passed the frozen candidate-superiority gate, and a later oracle analysis found no predeclared two-dataset headroom sufficient to justify a deployable selector; measured timings do not establish mobile energy use.Source: §5.6 · p. 10

RQ1

• HHAR’s original fold favors FedPer near saturation, but the complete nine-user result is lower with FedBN’s largest displayed mean; PAMAP2’s displayed maximum also changes, while paired leader gaps include zero.The authors interpret this as split sensitivity and descriptive method heterogeneity, not statistically established rank reversal; complete rotation or prospectively sampled participants belongs in the evaluation design.Source: §6 · p. 11

RQ2

• ATP-style leads WISDM gain and absolute score but has PAMAP2/SisFall losses; Feature-only has no seed-averaged 2-pp case but negative SisFall mean, while FedBN has one seed-averaged loss exceeding 5 pp.Feature-only also shows four 1-pp losses and harmful individual seed realizations; FedBN improves all five dataset means. The authors call for absolute baselines, a matched zero-shot branch, and explicit harm frequency, severity, uncertainty, and realization variability.Source: §6 · p. 11

RQ3

• Calibration can omit classes, be dominated by one activity, or be unavailable for short recordings; more windows do not guarantee improvement, and unlabeled deployment prevents the offline harm indicator from selecting the correct branch.The failed oracle headroom gate means a practical abstention rule requires independent, prospectively validated observable signals.Source: §6 · p. 11

RQ4

• High fall sensitivity can coexist with low specificity and repeated false-positive windows; grouped fall F1 alone can obscure weak activity discrimination, so UP-Fall remains a descriptive impact-phase stress case.The benchmark does not convert correlated windows into operational alarm episodes and retains multiclass quality and confusion counts alongside the fall view.Source: §6 · p. 11

What the benchmark contributes

• The benchmark’s contribution is an auditable onboarding validity contract spanning unseen-client FL, unlabeled calibration, recording-level separation, participant-level inference, branch-matched harm, tail quality, availability, and fall false-positive windows.Corrected HHAR history motivates artifact-level leakage checks, while the rotation result shows that even a leakage-free fold can be unrepresentative.Source: §6 · p. 11

Development rather than confirmation

• The HHAR/PAMAP2 rotations were predeclared after choices froze but remain part of development; effect estimates need independent replication before confirmatory or safety claims.The study reports no uncorrected multiple-comparison significance tests.Source: §7 · p. 11

Statistical scope

• Headline domains contain only 6–9 outer units, and three seeds reduce stochastic noise but do not add participants.Only six methods have expanded HHAR/PAMAP2 coverage; SisFall, UniMiB, and WISDM retain the frozen original split, so all 21 methods are not fully rotated. Overlapping source-training sets, jointly varying training and calibration composition, and higher-budget missingness limit interval, seed-audit, and cross-budget interpretations; leave-one-person-out sensitivity cannot replace a larger prospective sample.Source: §7 · p. 11

Method and model scope

• The six-strategy comparison uses one common TCN, source-training recipe, and one-step episodic onboarding; ATP-style and broad-screen proxies are not author-validated exact reproductions.FedBN uses BatchNorm while other core methods use GroupNorm, making its within-branch gain structurally different. Wearable-specific methods OFTTA, COA-HAR, SIGHT, and PI-TTA [22] [23] were discussed but not tested, so the tables are not a state-of-the-art wearable TTA leaderboard; other architectures or prospective controls could change conclusions.Source: §7 · pp. 11–12

Federated scope

• Federated clients are simulated from participant partitions, and the evidence supports a protocol for checkpoints from one federation recipe—not a causal claim that federation creates negative transfer.The centralized-source reference lacks complete HHAR/PAMAP2 rotations and cannot isolate whether split sensitivity is unique to federated source training.Source: §7 · p. 12

Dataset and deployment scope

• Controlled public datasets and simulated clients do not establish clinical effectiveness, real-device performance, or event-level fall detection.UP-Fall is a local five-subject skeleton derivative with one reported outer unit, not the full multimodal benchmark. Latency, battery use, streaming drift, and prospective user experience remain untested; window-level fall metrics do not estimate false alarms per hour or event-level detection, and archived labeled calibration streams do not support uniform post-hoc class-coverage analysis.Source: §7 · p. 12

Privacy, fairness, and safety

• The study makes no universal privacy, demographic-fairness, or deployment-safety claim.Keeping raw records local in simulation does not guarantee differential privacy, secure aggregation, or attack resistance; UniMiB age groups are not inferred from identifiers, and “material harm” is an analysis threshold rather than a medical risk threshold.Source: §7 · p. 12

8 Prospectively Scoped Future Work

• OFTTA should be evaluated on complete HHAR/PAMAP2 rotations under the recording-disjoint outer-person protocol, and frozen analyses and harm thresholds tested on an untouched cohort or dataset.The proposed work also includes porting author code for a wearable-specific method such as OFTTA.Source: §8 · p. 12
• Centralized and federated source checkpoints should be compared under identical outer splits, while independent calibration-recording/window draws separate training from calibration variance.Each fixed training checkpoint would be evaluated under multiple independently indexed calibration draws.Source: §8 · p. 12
• Event-clustered fall alarms, multiple backbones, streaming drift, and real-device energy are also prospective topics, not completed evidence.Future results must not erase the present negative outcomes.Source: §8 · p. 12

9 Reproducibility and Artifact Availability

• The package includes 567 broad jobs, 270 rotation jobs, 837 combined jobs, and 18 hashed derived files for exact table regeneration and coverage auditing.It also contains protocol YAML files, reporting scripts, job configurations and metrics, derived CSVs and figures, source-archive hashes, a hostile-review audit, and a final SHA-256 registry.Source: §9 · p. 12
• Canonical split files are absent, so the package does not claim from-scratch recreation of every source split.A submission release must include verified canonical manifests or explicitly preserve this limitation; hashes alone do not replace their contents.Source: §9 · p. 12
• Two hash-registered sensitivity packages derive robustness analyses from archived metrics without adding training runs or participants.They produce per-person maxima, harm-threshold counts, leave-one-person-out ranges, leader intervals, zero-event bounds, a seed-realization audit, and a participant heatmap; the second package verifies 684 primary seed–person–method rows.Source: §9 · p. 12

10 Conclusion

• Completing the HHAR rotation lowered every mean by 13.8–17.8 pp from the near-saturated scores on one frozen person.The displayed leader changes, but paired intervals leave the top methods as unresolved near-ties rather than proving a rank reversal.Source: §10 · p. 13
• No adaptive core mechanism combines positive mean gain across all datasets with zero observed 2-pp seed-averaged harm.ATP-style combines gains with multiple severe losses; Feature-only has lower seed-averaged severity but harmful individual realizations and no finite-sample safety guarantee; FedBN shows a stable positive mean can coexist with a severe loss and conditional absolute quality.Source: §10 · p. 13
• The displayed leader changes, but paired intervals leave the top methods as unresolved near-ties rather than proving a rank reversal.The complete HHAR rotation lowers every mean by 13.8–17.8 pp from the near-saturated scores on one frozen person.Source: §10 · p. 13
• Calibration quantity, lower-tail outcomes, and fall-window specificity expose further failures hidden by mean accuracy.Evaluation should preserve leakage controls, complete or prospectively sampled outer-person units, matched source checkpoints, separate participant and realization estimands, availability denominators, and negative results before deployment claims.Source: §10 · p. 13

Author contributions statement

• R.A., V.K.R., S.M., and T.S. contributed across methods, experiments, analysis, interpretation, supervision, and manuscript preparation or review; all authors approved the final manuscript.R.A. prepared the original draft; V.K.R. contributed to experimental design and validation; S.M. to critical review; and T.S. to supervision and research guidance.Source: Author contributions statement · p. 13

Conflict of Interest

• The authors report no financial or personal affiliations affecting the study and acknowledge no conflicts of interest.The same declaration appears in the additional information and conflict-of-interest passages.Source: Additional information · p. 13, Conflict of Interest · p. 13

Additional information

• Accession codes are not applicable.Source: Additional information · p. 13

Data availability

• All source code, configuration files, and supplementary scripts are publicly available; datasets remain under their original license.The supplied passage gives the repository as https://github.com/TitanRahil/When-Adaptation-Hurts.Source: Data availability · p. 13

Clinical Trial Number

• The clinical trial number is not applicable.Source: Clinical Trial Number · p. 13

Ethics, Consent to Participate, and Consent to Publish

• Ethics, consent to participate, and consent to publish declarations are not applicable.Source: Ethics, Consent to Participate, and… · p. 13

Author Biographies

• Rahil Aftab studies computer science at Jamia Hamdard and researches federated learning, privacy-preserving AI, computer vision, and trustworthy AI.He completed a research internship at VECC, where he worked on federated learning for privacy-preserving AI systems.Source: Author Biographies · p. 15
• Vineet Kumar Rakesh is a VECC Technical Officer with over 23 years of experience in software engineering, databases, and AI.His research includes talking-head generation, lip reading, and ultra-low-bitrate video compression; he is pursuing a Ph.D. at Homi Bhabha National Institute.Source: Author Biographies · p. 15
• Soumya Mazumdar pursues degrees in data science and computer science, with research on temporal generative modeling, computer vision, and controllable video synthesis.At VECC, he worked on pose- and landmark-conditioned video generation, benchmarking, and deployment pipelines; he has contributed to publications and an Indian patent.Source: Author Biographies · p. 15
• Dr. Tapas Samanta leads VECC’s Computer and Informatics Group, with expertise spanning AI, industrial automation, embedded systems, high-performance computing, and accelerator control.He also leads technology-transfer initiatives and public scientific outreach at VECC.Source: Author Biographies · p. 15