Whether wearable photoplethysmography (PPG) encodes dietary information remains unknown. We trained a language model on 1.1M meals to align PPG with meal descriptions. Near-meal PPG carries a meal-specific signal beyond individual traits; predictability decreases farther from meals. This transfers to dietary tasks: PPG increases AUC by 11% for a caloric-intake proxy and satiety across held-out and independent cohorts. Wearable PPG may augment text-based logging, compensating for brief descriptions.
Digital mental health interventions show promise but have yet to meet expectations. We discuss gaps between promising results of pilot studies and null findings in randomized controlled trials, focusing on two major approaches: just-in-time relapse prevention and just-in-time adaptive interventions. We propose that relapse prevention will profit from long-term studies capturing emerging episodes, and we suggest a novel framework for adaptive interventions including causality testing to proceed towards effective trialing.
Affordable electroencephalography (EEG) headsets could widen access to EEG-based treatment and assessment, but require rigorous validation. We evaluated PainWaive, a custom-developed 2-channel central scalp EEG headset designed for remote EEG monitoring and neurofeedback research in chronic pain. Eighty participants completed two resting-state EEGs using PainWaive and a research-grade system (LiveAmp). Alpha, beta and theta power, and peak alpha frequency (PAF), were derived from homologous central channels. PainWaive demonstrated good-to-excellent test-retest reliability for alpha power and PAF (ICCs = 0.72–0.93), with beta and theta ranging from moderate-good (ICCs = 0.42–0.76). Across 160 home-based neurofeedback sessions in two chronic pain samples, reliability was good-to-excellent across metrics (ICCs = 0.72–0.88). Spectral-shape correlations between PainWaive and LiveAmp were excellent (r > 0.91). Cross-device agreement was moderate-to-good for alpha power and PAF (ICCs = 0.53–0.74), but weaker for beta and theta (ICCs = 0.16–0.51). Sensitivity analyses showed alpha power and PAF were robust across choices, whereas beta and theta agreement varied by pipeline. Overall, PainWaive demonstrated reliable measurement of central scalp EEG metrics, supporting repeated within-device monitoring. Cross-device findings showed correspondence with research-grade EEG for some measures but should not be interpreted as establishing direct numerical interchangeability between systems. These findings provide a technical foundation for future remote EEG and neurofeedback studies in chronic pain.
Heart failure with preserved ejection fraction (HFpEF) complicated by unstable angina (UA) represents a major challenge in chronic disease management with high risk of recurrent heart failure hospitalization (HFH), and despite the availability of multicenter real-world data, AI-driven tools for multi-horizon risk assessment remain limited. This multicenter retrospective cohort enrolled 4367 patients with UA and HFpEF, and an ensemble feature selection pipeline integrating LASSO, SVM-RFE and Boruta identified robust predictors, followed by training of twelve machine learning algorithms to estimate HFH risk at 12-, 24- and 36-month horizons with external validation. The optimal time-specific models (ET, GBDT and LightGBM) achieved strong discrimination in the training cohort with AUCs of 0.818, 0.834 and 0.863, and maintained performance in external validation with AUCs of 0.839, 0.836 and 0.851, respectively, while SHAP analysis highlighted diabetes mellitus, TyG-BMI and NT-proBNP as the most influential predictors. We developed and externally validated a multi-time-point machine learning framework for HFH risk stratification in UA and HFpEF, which showed acceptable calibration and potential clinical utility on decision-curve analysis. Further prospective and real-world validation across diverse populations and healthcare systems is needed to confirm clinical effectiveness and support translation into routine practice.
Pregnancy dating fails the women who need it most: those who reach care without a first-trimester scan. We externally validated a commercial image-only ultrasound deep-learning model for delivery-date prediction, applied unmodified to 64,235 frames from 3788 examinations of 2550 pregnancies (34% migrant) at four Chilean public hospitals. Mean absolute error was 6.9 days versus 11.6 for the conventional due date, yet most of this gain reflected a shorter local gestation and fell below one day against recentred comparators. Fixed-term comparators assign every pregnancy the same gestational age at birth and therefore cannot rank who will deliver early, at any level of point accuracy: accuracy metrics are blind to the ranking on which management turns whenever the outcome distribution is concentrated, as in term-dominated cohorts. In a prespecified landmark analysis the model’s separate image-only risk score discriminated preterm birth with an area under the curve of 0.86, against 0.52 for gestational age alone; outside the landmark restriction it exceeded a clinical baseline of coded diagnoses plus same-examination biometry by 0.09, although it was miscalibrated. Error and calibration were also assessed by maternal nationality. These classification analyses are exploratory, and no woman was managed with the model; they identify what a prospective study should test.
