Respiratory-rate features add modest modality-related information beyond heart-rate features in heterogeneous wearable sport recordings: a secondary analysis

To test whether active-exercise features derived from respiratory rate (RR) add incremental modality-related information after accounting for heart rate (HR) features in heterogeneous wearable sport recordings. Observational secondary analysis of a public wearable exercise dataset. Of 126 source recordings, 124 recordings from 79 modality-qualified participant identifiers across 10 exercise modalities met the rule-based signal and annotation criteria. The primary feature set comprised four heart-rate and four respiratory-rate descriptors calculated on aligned active-exercise segments. Exercise modality was used as an analytical context label, not as a clinically useful classification endpoint. All primary models included the baseline covariate block B (age, sex, body mass index, and active-exercise duration). A fixed L2-penalised multinomial logistic regression was evaluated with 50 repeated modality-stratified identifier-grouped two-fold partitions. Primary inference used paired model differences, a modality-stratified identifier-level bootstrap of out-of-fold predictions, and an identifier-level paired permutation test. Mean identifier-level macro-F1 across repeated grouped partitions was 0.379 for B plus heart rate, 0.400 for B plus respiratory rate, and 0.475 for B plus both signals. Adding respiratory-rate features to heart-rate features changed macro-F1 by + 0.096 (95% bootstrap percentile interval + 0.061 to + 0.136; paired permutation p = 0.0045), whereas adding heart-rate features to respiratory-rate features changed macro-F1 by + 0.075 (95% bootstrap percentile interval + 0.031 to + 0.124). Expanding the HR reference from four to eight descriptors retained a ridge RR increment of + 0.108 (Holm-adjusted p = 0.0040); cross-model evidence remained limited. The random-forest estimate was positive (+ 0.050), but its 95% bootstrap percentile interval (+ 0.020 to + 0.085) and paired permutation test ( p = 0.093) provided discordant uncertainty evidence. Respiratory-rate features contributed modest modality-related information beyond matched heart-rate features within the primary ridge framework in this dataset. The findings describe context-sensitive wearable signal variation but do not establish test-retest reproducibility, measurement validity, training-load validity, external generalisability, or clinical utility.