Source-linked AI summary
DeepHeart: Semi-Supervised Sequence Learning for Cardiovascular Risk Prediction
Brandon Ballinger, Johnson Hsieh, Avesh Singh, Nimit Sohoni, Jack Wang, Geoffrey H. Tison, Gregory M. Marcus, Jose M. Sanchez, Carol Maguire, Jeffrey E. Olgin, Mark J. Pletcher
TL;DR
The paper addresses cardiovascular-risk prediction from noisy wearable data when medical labels are scarce. It trains a semi-supervised multi-task LSTM using large-scale unlabeled heart-rate measurements and compares two pretraining strategies. The reported results show high accuracy across four conditions and higher accuracy than biomarker-based baselines, while highlighting confounding and deployment as unresolved challenges.
Problem
Wearable data are abundant but noisy, variably sampled, and poorly supported by labeled medical databases, limiting supervised learning for health-condition prediction.
Method
DeepHeart combines a multi-task LSTM with semi-supervised sequence learning or heuristic pretraining using unlabeled wearable measurements and limited diagnosis labels.
Results
DeepHeart achieves high accuracy for four conditions and higher accuracy than biomarker-derived baselines for detecting diabetes, sleep apnea, hypertension, and high cholesterol.
Takeaways & Limitations
The findings suggest a possible public-health screening approach based on cardiovascular-risk signals from readily available, off-the-shelf heart-rate monitors.
Takeaways & Limitations
Prediction may reflect treatments or other confounders, such as beta blockers affecting heart-rate variability, and deployment faces costly errors, integration burdens, regulation, and misaligned incentives.
Abstract
from arXiv · showhide
We train and validate a semi-supervised, multi-task LSTM on 57,675 person-weeks of data from off-the-shelf wearable heart rate sensors, showing high accuracy at detecting multiple medical conditions, including diabetes (0.8451), high cholesterol (0.7441), high blood pressure (0.8086), and sleep apnea (0.8298). We compare two semi-supervised train- ing methods, semi-supervised sequence learning and heuristic pretraining, and show they outperform hand-engineered biomarkers from the medical literature. We believe our work suggests a new approach to patient risk stratification based on cardiovascular risk scores derived from popular wearables such as Fitbit, Apple Watch, or Android Wear.
Introduction
DeepHeart addresses the scarcity of labeled wearable data and the noise of ambulatory heart-rate measurements by combining large-scale unlabeled sensing with semi-supervised learning. The study reports higher accuracy than biomarker-based baselines for detecting four health conditions.
- Motivation: Wearable sensors generate abundant unlabeled measurements, but their errors, variable sampling, and activity-related confounders complicate medical prediction.Relevant activities include walking, exercise, stress, alcohol consumption, and coffee drinking.
- Motivation: Purely supervised learning is constrained because wearable databases with sufficient medical labels do not exist and recruiting labeled patients is costly.The paper contrasts scarce diagnosis labels with large volumes of wearable data.
- Approach: DeepHeart recruits 14,011 participants and uses 200 million unlabeled measurements for two semi-supervised training procedures.The procedures are semi-supervised sequence learning and heuristic pretraining.
- Approach: Semi-supervised sequence learning initializes supervised training from an LSTM sequence autoencoder, while heuristic pretraining learns heart-rate-derived biomarkers from medical literature.Both approaches use limited labeled data after pretraining.
- Results: DeepHeart achieves higher accuracy than strong biomarker-derived baselines when detecting diabetes, sleep apnea, high blood pressure, and high cholesterol.The introduction presents this as the paper’s main experimental comparison.
Data and Study Cohort
The study combines continuous wearable sensing with medical-history labels to build a multi-condition cohort for cardiovascular-risk prediction. Data are organized into user-week tensors with separate sensor channels, transformed time gaps, and masked diagnosis outputs.
- Study Cohort: The cohort includes 14,011 Apple Watch app users whose heart rate, steps, activity data, diagnoses, blood tests, and medications were collected through an IRB-approved study.Recruitment occurred through a partnership with UCSF’s Health eHeart Study.
- Study Cohort: The study targets high cholesterol, hypertension, sleep apnea, and diabetes because they are prevalent, often undiagnosed, and associated with cardiovascular risk.The passage reports substantial undiagnosed disease burden and mortality contributions for several conditions.
- Sensor Data: Wearables provide heart-rate and step-count channels at variable sampling intervals, including five-second workout measurements and roughly five-minute measurements otherwise.The data therefore span multiple temporal resolutions.
- Sensor Data: Measurements are encoded as user-week, timestep, and channel tensors, with heart rate and step count separated and intra-channel time gaps represented by a special dt channel.The representation accommodates different measurement timescales.
- Data Processing: After filtering incomplete weeks, the dataset contains 57,675 person-weeks split into training, tuning, and validation sets.Participant assignments are disjoint across the three sets.
- Labeled Outputs: Outputs contain separate masked diagnosis channels for high cholesterol, hypertension, sleep apnea, diabetes, and ECG abnormalities aligned to sensor timelines.Only timesteps with available task labels contribute to scoring.
Model Architecture
DeepHeart processes multi-channel, multi-timescale wearable measurements with temporal convolutions followed by bidirectional LSTMs. Its output layer produces a separate prediction for each task and timestep.
- Architecture: DeepHeart takes multi-channel, multi-timescale sensor measurements and produces multi-task, multi-timescale diagnosis predictions.The architecture is summarized in Figure 2.
- Convolutional Layers: Three temporal-convolution layers extract features and temporal invariances before recurrent processing.The first convolution uses a wide filter, while later layers use residual units, dropout, and max pooling.
- Recurrent Layers: Four bidirectional LSTM layers model longer temporal patterns in wearable signals and their corresponding diagnoses.Each layer has 128 units, split evenly between forward and backward directions.
- Output Layer: A filter-length-one convolution maps the final LSTM representation to one prediction per task and timestep.The final layer uses tanh, constraining outputs to the range -1 to 1.
Training Methods
The training study compares ordinary supervised learning with two pretraining strategies designed to exploit abundant unlabeled wearable data. One reconstructs sequences, while the other predicts heart-rate-variability heuristics.
- Training Strategy: Semi-supervised training is motivated by abundant unlabeled wearable data and scarce medical-condition labels.The comparison includes no pretraining, heuristic pretraining, and semi-supervised sequence learning.
- Semi-Supervised Sequence Learning: Semi-supervised sequence learning pretrains a sequence-to-sequence autoencoder and transfers its convolutional and recurrent weights to supervised training.The autoencoder uses the same three convolutional and four recurrent layers as the supervised architecture.
- Heuristic Pretraining: Heuristic pretraining predicts average absolute differences between successive heart-rate measurements over 5-second, 30-second, 5-minute, and 30-minute windows.These four outputs form a weakly supervised pretraining task inspired by beat-to-beat variability analysis.
Experiments
DeepHeart experiments compare supervised and pretrained LSTM variants with hand-engineered baselines across four cardiovascular-risk conditions. Semi-supervised sequence learning improves performance on three conditions, while label efficiency and input-channel analyses show condition-specific effects.
- LSTM without Pretraining: 0.80–0.85 c-statistics from the supervised LSTM exceeded hand-engineered baselines across sleep apnea, hypertension, high cholesterol, and diabetes.The reported comparisons were 0.80 vs 0.55–0.66, 0.80 vs 0.61–0.72, 0.67 vs 0.43–0.60, and 0.85 vs 0.41–0.79, respectively.
- Heuristic Pretrained LSTM: Heuristic pretraining raised high-cholesterol c-statistic from 0.67 to 0.71 but reduced high-blood-pressure performance from 0.80 to 0.74.Sleep apnea and diabetes showed neutral results, and the authors suggest the heuristic may encode useful nonlinear patterns for cholesterol but poor predictors for blood pressure.
- Semi-supervised Sequence Learning: Semi-supervised sequence learning significantly improved detection of high cholesterol, high blood pressure, and sleep apnea.The method pretrained DeepHeart as a sequence autoencoder before supervised training.
- Semi-supervised Sequence Learning: 0.83 c-statistic for sleep apnea was achieved despite few participants sleeping with their watch, suggesting daytime wearable signals capture a distinct apnea signature.The result is reported alongside ROC analyses comparing pretrained methods with the strongest overall baseline.
- Effect of Varying Amounts of Labeled Data: 10x better data-efficiency occurred for hypertension and sleep apnea, reaching nearly full-data accuracy with 10% of labeled data under unsupervised pretraining.The effect was muted for diabetes and high cholesterol; diabetes had only 1,678 labels, while high cholesterol showed little accuracy-label correlation.
- Ablative Analysis of Input Channels: Heart-rate-only inputs reduced diabetes and high-blood-pressure AUC by 0.0653 and 0.0721, indicating activity response is more informative than heart rate alone.Aggregate step counts achieved AUCs of 0.7011, 0.5811, 0.5714, and 0.5601 for diabetes, sleep apnea, hypertension, and high cholesterol.
Discussion
The paper presents semi-supervised wearable-based risk prediction as a promising response to scarce labeled medical data, while identifying confounding, deployment, interpretability, and data-scale challenges that remain before real-world use.
- Discussion: Semi-supervised sequence learning and heuristic pretraining address limited labeled data while detecting four common conditions with off-the-shelf heart-rate monitors.The authors frame this as a first step toward applying deep-learning techniques to medical screening.
- Discussion: Confounding may cause accurate hypertension predictions to reflect both disease and treatments such as beta blockers, motivating confounder-aware neural networks.The paper distinguishes predictive accuracy from identifying disease-specific causes.
- Discussion: Hospital deployment is constrained by costly errors, technical integration, regulatory burden, physician training, and misaligned financial incentives.The authors propose ambulatory mobile screening as an outside-in alternative that guides high-risk participants toward care.
- Discussion: Deep-learning systems remain difficult to interpret because their black-box reasoning and potential failure modes are not readily exposed to physicians or regulators.Differentiable attention is proposed as one mechanism for improving interpretability.
- Discussion: Long wearable records can contain nearly one million time steps per participant, motivating architectures that better model long-range physiological changes.The paper names Clockwork RNNs, Phased LSTMs, and Gaussian Process RNNs as possible approaches.