Recovery Latency (RL)—the time required for perceived recovery to return to individual baseline following stress—has recently been validated as a dynamic biomarker of adaptive capacity (3). However, environmental determinants of RL variation remain poorly characterized. At northern latitudes, photoperiod varies dramatically (~10 hours annually), providing a natural experiment to examine environmental modulation of recovery kinetics. We analyzed 11,915 recovery events from 73 competitive endurance runners monitored over 7.3 years (2012–2019) at 53.2°N latitude (Groningen, Netherlands). RL was defined as the number of days required for perceived recovery to return to ≥95% of individual baseline following high-load training events (top 25% of historical exertion). Daily photoperiod (7.5–17.0 h) was calculated from geographic coordinates. Mixed-effects models with athlete-specific random slopes tested linear and quadratic associations, adjusted for training load. A significant quadratic association revealed a U-shaped relationship between photoperiod and RL (β_quadratic=0.006, p=0.004). Optimal recovery occurred at 11.6 h photoperiod (95% CI: 9.9–13.3), corresponding to spring/autumn conditions. RL increased during winter (+4.8% from optimal) and summer (+8.5% from optimal). Substantial inter-individual variability in photoperiod sensitivity was observed (random slope variance=0.004, p<0.001). Load adjustment minimally altered photoperiod coefficients (Δβ=4.9%), suggesting seasonal training patterns do not fully explain the association. These findings demonstrate that recovery kinetics are optimized at intermediate photoperiod and impaired at seasonal extremes, supporting the view that adaptive capacity is modulated by environmental light exposure consistent with circadian-mediated pathways.

Two athletes complete identical training sessions. One recovers in 48 hours; the other requires 7 days. Standard monitoring shows similar HRV and training load — yet their adaptive capacities are fundamentally different. Current monitoring relies on static snapshots (HRV, load, biomarkers) that fail to capture this temporal dimension. Recovery Latency (RL) is defined here as the number of days required for a physiological marker to return to individual baseline following a high-load event — a time-based metric that may represent a more direct measure of adaptive capacity than any single static measurement. We analyzed 27 independent datasets from Kaggle public repositories (~230,000 observations; 23 unique after deduplication) to validate RL empirically across performance, injury, and clinical contexts. Objective To validate Recovery Latency as a physiological marker of adaptive capacity and to operationalize it within a latent-state framework — the Dynamic Adaptive Capacity (DAC) model — across multiple independent datasets comprising performance, injury, and autonomic recovery domains. Methods A total of 23 unique datasets (~230,000 observations) were analyzed, including athletic performance cohorts, injury surveillance systems, and HRV-based recovery studies. RL was calculated as the time required for recovery proxies to return to individual baseline following high-load events. The DAC model integrated RL dynamics into a continuous 0–100 individualized adaptive-capacity score with state classification (green/yellow/red). Associations with autonomic, endocrine, behavioral, and performance variables were tested using correlation analyses and predictive modeling. Results RL strongly predicted overtraining (AUC = 0.934), showing robust associations with cortisol (r = 0.631) and HRV (r = −0.644). In longitudinal running data, 28,472 RL events were identified across 74 athletes (~385 events/athlete), with 60% of athletes exhibiting a positive RL slope, indicating progressive loss of recovery efficiency. HRV datasets confirmed the physiological coherence of the model, differentiating clinical states (sepsis vs. control, p < 0.0001), tracking psychological stress, and demonstrating autonomic coupling with training load. Behavioral factors also contributed to recovery dynamics, with screen time inversely associated with mental wellness (r = −0.636, p < 0.001) and sleep quality. DAC provided an interpretable adaptive-state score, enabling individualized trajectory tracking and early detection of maladaptive trends. Conclusion. Recovery Latency is a valid and scalable biomarker that captures the temporal cost of stress and the efficiency of physiological recovery, providing discriminative validity beyond static markers in detecting maladaptation. When embedded in the DAC framework, it enables continuous quantification of adaptive capacity and actionable state classification. This approach supports a paradigm shift from load monitoring to recovery-dynamics monitoring, with direct applications in performance optimization, injury prevention, and real-time physiological decision-making.

Purpose: Respiratory-event rates quantify event frequency but do not describe how respiratory-related acoustic activity is expressed across an overnight recording. We investigated whether a score-defined digital acoustic representation captures information not reducible to annotated respiratory-event rate.Methods: We analyzed 32 overnight home respiratory polygraphy recordings from the APSAA dataset, including audio, respiratory annotations, and oxygen saturation. Twenty-eight recordings yielded at least one qualifying acoustic episode (8,512 episodes). Five acoustic features—episode density, relative intensity, duration, long-event occurrence, and temporal distribution—formed a composite score defining three acoustic phenotypes using fixed heuristic thresholds. Temporal compaction quantified short inter-episode intervals and was adjusted for episode quantity.Results: Scores were computationally reproduced with identical phenotype assignment in 28/28 recordings (mean absolute difference 0.0011; maximum 0.0046). Two recordings with comparable annotation-derived respiratory-event rates (30.45 vs 33.38 events/h; 217 vs 232 annotated events) and mean SpO₂ (91.75% vs 91.61%) contained 382 vs 5 acoustic episodes and scores of 0.685 vs 0.145. Load-adjusted temporal compaction differed among phenotypes (H=8.192, p=0.0166, ε²=0.248). The effect persisted after removing the historical 120-s interval exclusion (H=8.108, p=0.0174, ε²=0.244); residual association with episode count remained. The low- versus intermediate-score contrast was robust, whereas the high-score contrast was less stable.Conclusions: Overnight respiratory audio can provide reproducible digital acoustic phenotypes describing the amount and temporal organization of detected acoustic activity beyond annotated event frequency. These are signal representations, not clinical phenotypes or sleep stages. Repeated-night studies are required to establish within-person stability, transitions, and clinical meaning.

BACKGROUND The proliferation of artificial intelligence (AI) models in health has been accompanied by a fundamental but largely unexamined assumption: that the availability of health data and computational capacity is, in itself, sufficient justification for training new models. This assumption has driven an exponential increase in publications reporting AI applications in clinical and public health domains, yet the proportion of these models that demonstrably improve patient outcomes, clinical decision-making, or health system efficiency remains remarkably low. The field lacks a systematic pre-training evaluation framework that distinguishes between what is technically trainable and what is scientifically and socially necessary. OBJECTIVE This paper proposes a methodological framework for determining when training an AI model in health is scientifically justified and when it is not. We argue that the decision to train should be treated as a consequential resource allocation decision, subject to the same standards of justification applied to any health intervention, rather than as a neutral technical exercise. METHODS We conducted a conceptual and integrative review of the current AI-in-health literature, examining prevailing training practices, outcome reporting standards, and the gap between predictive performance and clinical impact. Drawing on frameworks from health technology assessment, implementation science, and philosophy of science, we synthesized a set of pre-training criteria organized around four dimensions: scientific necessity, clinical relevance, economic justification, and social accountability. RESULTS We present the TRAIN-H (Training Rationale Assessment for AI in Health) framework, a structured decision tool comprising seven core questions that must be affirmatively answered before model development is justified. The framework formalizes the principle that prediction without consequence is not impact, and that accuracy without altered conduct, workflow, or cost does not constitute a valid reason to build a model. We identify six explicit conditions under which training should not proceed and discuss implications for research ethics, peer review, funding allocation, and graduate education. CONCLUSIONS Training AI in health is an investment of computational, human, institutional, and public resources. Like any health intervention, it requires a clear hypothesis, a defined population benefit, and a measurable endpoint. The absence of these elements does not merely weaken a study—it eliminates the justification for the model’s existence. We call for the adoption of pre-training justification standards across research institutions, funding bodies, and editorial boards.

Background Recovery Latency (RL) is a longitudinal metric capturing the time required for physiological variables to return to baseline following training or competitive loads. Prior work in this series demonstrated that RL reflects adaptive states across large longitudinal athletic datasets. However, whether the temporal dynamics of RL carry information about the stability of the underlying physiological system — beyond mean recovery levels — remains unknown. Theory from complex adaptive systems predicts that systems approaching a critical transition exhibit statistical precursors, notably increasing variance and lag-1 autocorrelation, collectively referred to as Early Warning Signals (EWS). Objective To determine whether temporal dynamics of Recovery Latency — specifically rolling variance and rolling autocorrelation — exhibit patterns consistent with Early Warning Signals of declining physiological resilience in longitudinal athlete datasets. Methods We analysed 28,472 recovery events from 649 athletes across three independent datasets within the RL series dataset (recuperação_VFC_100k, athlete_physiological, lesões_multimodal). For each athlete, rolling variance (EWS₁) and rolling lag-1 autocorrelation (EWS₂) of RL were computed using a 14-event window. Trend analysis was performed using Kendall τ statistics, complemented by sliding-window Kendall τ (window=20), flickering dynamics (state-transition rate), surrogate testing (n=100 permutations), and a Composite Early Warning Index (CEWI). Athletes were stratified into Declining (RL slope > 0.015 d/event, n=155) and Stable (RL slope < −0.010 d/event, n=131) adaptive states. An independent screen-time dataset (N=400) was analysed to examine environmental circadian perturbation as a potential driver of resilience loss. Results Rolling variance of RL increased significantly in athletes with declining adaptive capacity compared with stable athletes across all datasets (VFC_100k: Δτ=+0.516, p<0.001; Athlete Physiol.: Δτ=+0.556, p=0.001; Lesões Multimodal: Δτ=+0.281, p<0.001). Sliding-window Kendall τ confirmed temporal divergence between groups in the final observation window (τ_last combined: Declining=−0.009 vs Stable=−0.284, p=0.005). Rolling autocorrelation (p=0.052–0.922) and flickering dynamics (p=0.261–0.985) did not significantly differentiate groups. Screen exposure was strongly associated with a RL proxy index (ρ=+0.620, p<0.001), mediated through sleep quality degradation and stress elevation, supporting a circadian disruption pathway to resilience loss. Conclusions Temporal dynamics of Recovery Latency contain detectable Early Warning Signals of physiological resilience loss, primarily expressed as increasing rolling variance in athletes with deteriorating adaptive capacity. Replication across three independent datasets supports RL variability as a scalable marker of adaptive instability. Associations with screen exposure further suggest that environmental circadian perturbations may contribute to resilience decline, linking chronobiology, behavioral factors, and recovery physiology.

Objective : To examine Recovery Latency (RL) as a candidate temporal marker of adaptive capacity by assessing whether recovery dynamics can be consistently quantified across independent sport and physiological data sets. Methods : A total of 23 unique data sets (∼230,000 observations) were analyzed including athletic performance cohorts, injury surveillance systems, and heart rate variability-based recovery studies. RL was defined as the time required for recovery proxies to return to an individualized rolling baseline following high-load events. Analyses focused on sport-performance contexts, with additional data sets included as exploratory extensions. Associations with autonomic and performance-related variables were evaluated using correlation analyses and predictive modeling. The Dynamic Adaptive Capacity framework was explored as a potential application of RL dynamics. Results : Preliminary evidence suggests that RL captures variation in recovery dynamics. RL showed high apparent discriminative performance for overtraining risk (area under the receiver operating characteristic curve = 0.934) and was associated with physiological markers including cortisol ( r = .631) and heart rate variability ( r = −.644). Longitudinal analysis identified 28,472 RL events across 74 athletes, with 60% exhibiting a positive RL slope suggesting progressive changes in recovery dynamics across individuals. Patterns were consistent with expected physiological responses to training load and stress. Conclusion : These findings suggest that RL may be relevant in sport-performance contexts; however, they should be interpreted as exploratory and hypothesis-generating rather than as evidence of validation.