Adding wearable heart-rate data to tree-based models improved mortality prediction in heart failure patients with tricuspid regurgitation, with AUPRC increasing from 0.72 to 0.78.
Does adding wearable heart rate data to clinical models improve the prediction of 2-year all-cause mortality in heart failure patients with tricuspid regurgitation undergoing TTVI?
Adding continuous wearable heart rate data to clinical models moderately improves the prediction of 2-year all-cause mortality in heart failure patients undergoing transcatheter tricuspid valve intervention.
Absolute Event Rate: 0% vs 0%
Abstract Background Machine learning based predictive models for tricuspid regurgitation (TR) have largely focused on ‘static’ clinical, laboratory and imaging data. Yet dynamic physiological signals captured by consumer wearables - especially continuous heart rate streams - remain largely untapped in this population. Purpose We aimed to investigate whether adding wearable heart rate data into tree-based models improves prediction of mortality in heart failure patients with TR. Methods 479 patients who underwent Transcatheter Tricuspid Valve Intervention (TTVI) between 2017 and 2023 were enrolled in this study. Patients were provided with wearable Fitbit devices prior to TTVI. Hourly heart rate data were extracted, after which harmonic modelling and dimensionality reduction were applied to create multiple data sets. 2-year all-cause mortality was used as the primary outcome. Tree-based machine learning algorithms- Extreme Gradient Boosting (XGBoost), Random Forest (RF) and Light Gradient Boosting Machine (LightGBM)- were used to predict mortality. AUPRC, precision, recall/sensitivity, specificity and weighted F1 scores were used to assess the performance of the models. Results The median age for patients in the study was 80 (IQR 76-83) years and 50.7% (n=243) were females. All-cause mortality was observed in 38% of the patient cohort with a median follow-up time of 411 (IQR 168-1039) days. RF of Clinical + Latent Heart-Rate Embeddings (C-LHRE) was the best performing model with AUPRC, precision, recall, and specificity of 0.78, 0.77, 0.74, and 0.87 respectively. The best performing clinical only model was the RF with 0.72, 0.68, 0.83, and 0.78 for AUPRC, precision, recall, and specificity respectively. C-LHRE out-performed the clinical-only model (ΔAP Average Precision = 0.056; 95 % CI, -0.010 to 0.123; bootstrap p = 0.047). At the operating threshold that maximized the weighted F1 score on the test set (probability ≥ 0.41), the C-LHRE model achieved a weighted F1 score of 0.82, compared to 0.77 for the clinical-only model at the same threshold. SHAP analysis identified dimensionally reduced representations of minimum, maximum, and overall heart rate among the top 15 features. Conclusion In this study, adding wearable heart-rate data to tree-based predictive models for mortality moderately improved performance highlighting the potential relevance of wearable data. The observed shifts suggest that adding wearable heart-rate data helps the model to flag fewer false positives and deliver more actionable recommendations like the introduction of new therapy. However, this modest gain still demands multi-centre, multi-device validation and prospective impact studies to confirm generalisability and clinical utility before routine adoption. Given that wearables stream heart-rate data passively and at scale, their inclusion offers the possibility of a high-yield, low-effort path to improving risk stratification and individual patient treatment.Precision-Recall Curve
Idakwo et al. (Thu,) reported a other. Adding wearable heart-rate data to tree-based models improved mortality prediction in heart failure patients with tricuspid regurgitation, with AUPRC increasing from 0.72 to 0.78.