CoChat
Sign Up
Cited by 146Artificial Intelligence in HealthcareNon-Invasive Vital Sign MonitoringMachine Learning in Healthcare

DeepHeart: Semi-Supervised Sequence Learning for Cardiovascular Risk Prediction

Open access
Ballinger, Brandon, Hsieh, Johnson, Singh, Avesh, Sohoni, Nimit, Wang, Jack, Tison, Geoffrey, Marcus, Gregory, Sanchez, Jose +3 more·2018·Proceedings of the AAAI Conference on Artificial Intelligence
DOI 10.1609/aaai.v32i1.11891
Open this paper in CoChat — free account

Abstract

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 training 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.

AI snapshot

Written from the paper’s full text.

TL;DR

DeepHeart trains a semi-supervised, multi-task LSTM on wearable heart-rate and step-count data from 14,011 participants, detecting diabetes, high cholesterol, high blood pressure, and sleep apnea with c-statistics up to 0.8451. Semi-supervised sequence learning and heuristic pretraining outperform hand-engineered biomarker baselines, though the paper notes confounding, deployment, interpretability, and data-scale challenges.

Key findings

  • On 57,675 person-weeks, semi-supervised sequence pretraining achieved c-statistics of 0.7441 for high cholesterol, 0.8086 for high blood pressure, and 0.8298 for sleep apnea; the abstract reports diabetes at 0.8451, while Table 2 lists 0.8451 for the no-pretraining LSTM and 0.7998 for unsupervised pretraining.
  • The LSTM without pretraining outperformed baselines on sleep apnea (0.80 vs 0.55–0.66), hypertension (0.80 vs 0.61–0.72), high cholesterol (0.67 vs 0.43–0.60), and diabetes (0.85 vs 0.41–0.79).
  • Heuristic pretraining improved high cholesterol (0.71 vs 0.67) but significantly hurt high blood pressure (0.74 vs 0.80); semi-supervised sequence learning improved high cholesterol, high blood pressure, and sleep apnea.
  • For hypertension and sleep apnea, unsupervised pretraining yielded a 10x improvement in data-efficiency: nearly the same accuracy with 10% of labeled data as with 100% and no pretraining.

Methods

  • Recruited 14,011 Apple Watch app users in partnership with the UCSF Health eHeart Study, collecting 200 million unlabeled sensor measurements; after filtering, 57,675 person-weeks were split into 33,628 training, 18,555 tuning, and 12,790 validation weeks, with disjoint participants across sets.
  • Model inputs were multi-channel, multi-timescale tensors for heart rate, step count, and time deltas; architecture used 3 temporal convolution layers followed by 4 bidirectional LSTM layers, with multi-task outputs for high cholesterol, hypertension, sleep apnea, diabetes, and ECG.
  • Compared supervised LSTM, heuristic pretraining on heart-rate-variability metrics at 5-second, 30-second, 5-minute, and 30-minute windows, and semi-supervised sequence autoencoder pretraining; baselines used 13 hand-engineered biomarkers and standard ML algorithms, evaluated by c-statistic/AUC on the test set.

Limitations

  • The paper states challenges including confounding (e.g., beta blockers affecting heart rate variability), deployment barriers, lack of interpretability in deep learning, and difficulty modeling long time-scale and high-volume data.
  • The paper notes high cholesterol has the least direct physiological relationship with heart rate variability and may reflect confounders such as age, sex, or medications; diabetes has scarce labels (1,678 person-weeks).
  • Evident from the text: consumer sensors have significant error and variable measurement rates, ambulatory activities like walking, exercise, stress, alcohol, or coffee may confuse simple heuristics, and the results are predictive rather than causal.

Model-written summary. Check it against the paper before relying on it.

Integrity signals

Registry

Crossref · 10.1609/aaai.v32i1.11891

Cites

40 of 40 references verified0 citation errors · 0 not checkable

Statistics

Nothing to recompute.

Claims the paper makes

3 of 16 cited sources checked · 2 supported · 0 flagged · accuracy not yet measured

  • Premisecitation
    § Unsupervised Sequence Pretraining · p. 3

    Semi-supervised sequence learning has been shown to improve performance when presented with large amounts of unlabeled training data.

    “Semi-supervised sequence learning has been shown to improve performance when presented with large amounts of unlabeled training data (Dai and Le 2015).”
    Quote in the paperSubstantial claim1 of 1 sources checked
    • cites Dai and Le 2015 source
  • Findingown data
    § Abstract · p. 1

    Semi-supervised sequence learning and heuristic pretraining outperform hand-engineered biomarkers from the medical literature.

    “We compare two semi-supervised training methods, semi-supervised sequence learning and heuristic pretraining, and show they outperform hand-engineered biomarkers from the medical literature.”
    • Also stated in: § Introduction · p. 1 · plain
      “Then we present experimental results showing that semisupervised deep neural networks achieve higher accuracy in detecting diabetes, sleep apnea, high blood pressure, and high cholesterol than several strong baselines derived from biomarkers in the medical literature.”
    Quote in the paperSubstantial claim
  • Findingown data
    § Abstract · p. 1

    DeepHeart achieved a c-statistic of 0.7441 for detecting high cholesterol.

    “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).”
    Quote in the paperSubstantial claim
  • Findingown data
    § Heuristic Pretrained LSTM · p. 4

    Heuristic pretraining significantly worsened high blood pressure from 0.80 to 0.74.

    “When the supervised LSTM is initialized with weights from heuristic pretraining, we saw noticeable improvements for high cholesterol (0.71 vs 0.67), neutral results for sleep apnea and diabetes, and a significant loss for high blood pressure (0.74 vs 0.80).”
    Quote in the paperSubstantial claim
  • Findingown data
    § Discussion · p. 5

    DeepHeart shows high accuracy at detecting four common health conditions-diabetes, sleep apnea, hypertension, and high cholesterol-using readily available, off-the-shelf heart rate monitors.

    “Furthermore, we've shown high accuracy at detecting four common health conditions-diabetes, sleep apnea, hypertension, and high cholesterol-using readily available, off-the-shelf heart rate monitors.”
    Quote in the paperSubstantial claim
  • Classified as probably not a substantial claim. 10
  • May combine two claims. 12

  • Classified as probably not a substantial claim.

    Premisecitation
    § Weakly-Supervised Heuristic Pretraining · p. 3

    Previous work has applied statistical methods on beat-to-beat variability to detect heart arrhythmias.

    “Previous work has applied statistical methods on beat-tobeat variability to detect heart arrhythmias (McManus et al. 2013).”
    Quote in the paperSubstantial claim1 of 1 sources checked
    • cites McManus et al. 2013 source

  • Probably not a substantial claim: may not pass S3 (load-bearing).

    Premisecitation
    § Introduction · p. 1

    Deep neural networks have shown high accuracy at pattern recognition from noisy, complex inputs, including automated detection of diabetic retinopathy from images, skin cancer from mobile phone cameras, and onset of health conditions from electronic medical records.

    “Deep neural networks (LeCun, Bengio, and Hinton 2015) have shown high accuracy at pattern recognition from noisy, complex inputs, including automated detection of diabetic retinopathy from images (Gulshan et al. 2016), skin cancer from mobile phone cameras (Esteva et al. 2017), and the onset of health conditions from electronic medical records (Razavian, Marcus, and Sontag 2016; Lipton et al. 2015”
    Quote in the paperNot a substantial claim1 of 3 sources checked
    • cites LeCun, Bengio, and Hinton 2015 source
    • cites Gulshan et al. 2016 source
    • cites Esteva et al. 2017 source

      Split into 2 atoms — 0 confirmed, 2 partially supported, 0 disputed, 0 misattributed, 0 unverifiable. Expand the claim to see per-atom verdicts.

      Experimental — accuracy not yet measured

  • May combine two claims.

    Findingown data
    § Semi-supervised Sequence Learning · p. 4

    Semi-supervised sequence learning resulted in a significant improvement on high cholesterol, high blood pressure, and sleep apnea.

    “This approach resulted in a significant improvement on high cholesterol, high blood pressure, and sleep apnea.”
    Quote in the paperSubstantial claim

  • May combine two claims.

    Findingown data
    § Semi-supervised Sequence Learning · p. 4

    High cholesterol has the lowest performing prediction out of the four disease states, with a c-statistic of 0.74.

    “While we are able to detect high cholesterol with c-statistic of 0.74, it is the lowest performing prediction out of the four disease states, suggesting that much of the correlation may be a reflection of confounding variables such as age, sex, or usage of medications like beta blockers.”
    Quote in the paperSubstantial claim

  • May combine two claims.

    Findingown data
    § Pre-Processing · p. 2

    After filtering, there were 57,675 person-weeks of data in total, divided into 33,628 for training, 18,555 for tuning, and 12,790 for validation.

    “After filtering, there were 57,675 person-weeks of data in total, divided into 33,628 for training, 18,555 for tuning, and 12,790 for validation.”
    Quote in the paperSubstantial claim

  • May combine two claims.

    Findingown data
    § Effect of Varying Amounts of Labeled Data · p. 4

    For hypertension and sleep apnea, unsupervised pretraining yields a 10x improvement in data-efficiency: the DNN achieves nearly the same accuracy with 10% of labeled data using unsupervised pretraining as it does with 100% of labeled data using no pretraining.

    “For hypertension and sleep apnea, unsupervised pretraining yields a 10x improvement in data-efficiency: the DNN achieves nearly the same accuracy with 10% of labeled data using unsupervised pretraining as it does with 100% of labeled data using no pretraining.”
    Quote in the paperSubstantial claim

  • May combine two claims.

    Findingown data
    § Ablative Analysis of Input Channels · p. 5

    A model using only aggregate step counts achieved an AUC of 0.7011 for diabetes, 0.5811 for sleep apnea, 0.5714 for high blood pressure, and 0.5601 for high cholesterol.

    “A model using only aggregate step counts achieved an AUC of 0.7011 (diabetes), 0.5811 (sleep apnea), 0.5714 (hbp), and 0.5601 (high cholesterol), suggesting that although overall physical activity plays an important role in predicting the onset of disease, the heart's response to physical activity is a more salient biomarker that can be captured using deep learning.”
    Quote in the paperSubstantial claim

  • Probably not a substantial claim: may not pass S3 (load-bearing).

    Premisecitation
    § LSTM without Pretraining · p. 4

    Diabetes may impact heart rate variability through effects on the autonomic nervous system.

    “Previous work from the medical community has suggested that diabetes may impact heart rate variability through effects on the autonomic nervous system (Kudat et al. 2006).”
    Quote in the paperNot a substantial claim0 of 1 sources checked
    • cites Kudat et al. 2006 source

  • Classified as probably not a substantial claim.

    Study design and conductnone given
    § Pre-Processing · p. 2

    Each participant was randomly assigned to either the training, tuning, or testing set.

    “Each participant was randomly assigned to either the training, tuning, or testing set.”
    Quote in the paperSubstantial claim

  • May combine two claims.

    Interpretationown data
    § Discussion · p. 5

    Two methods, semi-supervised sequence learning and heuristic pretraining, address a key technical challenge in applying deep learning to medicine: achieving high accuracy with limited labeled data.

    “We've shown that two methods, semi-supervised sequence learning and heuristic pretraining, address a key technical challenge in applying deep learning to medicine: achieving high accuracy with limited labeled data.”
    Quote in the paperSubstantial claim

  • Probably not a substantial claim: may not pass S3 (load-bearing). May combine two claims.

    Premisecitation
    § Study Cohort · p. 2

    Sleep apnea is estimated to affect 22 million people in the US alone, over 80% of whom are undiagnosed.

    “Sleep apnea is estimated to affect 22 million people in the US alone, over 80% of whom are undiagnosed (Young et al. 1997).”
    Quote in the paperNot a substantial claim0 of 1 sources checked
    • cites Young et al. 1997 source

  • Probably not a substantial claim: may not pass S3 (load-bearing).

    Premisecitation
    § Study Cohort · p. 2

    High blood pressure and high cholesterol together account for 17.3% of global deaths.

    “Together, high blood pressure and high cholesterol account for 17.3% of global deaths (Alwan and others 2011), as they are both major risk factors for heart attacks.”
    Quote in the paperNot a substantial claim0 of 1 sources checked
    • cites Alwan and others 2011

      Couldn’t be matched to a paper in Crossref, so it can’t be checked.

  • Probably not a substantial claim: may not pass S3 (load-bearing).

    Premisecitation
    § Study Cohort · p. 2

    45.8% of diabetes cases are undiagnosed.

    “45.8% of diabetes cases are undiagnosed (Beagley et al. 2014).”
    Quote in the paperNot a substantial claim0 of 1 sources checked
    • cites Beagley et al. 2014 source

  • May combine two claims.

    Interpretationown data
    § Ablative Analysis of Input Channels · p. 5

    Although overall physical activity plays an important role in predicting the onset of disease, the heart's response to physical activity is a more salient biomarker that can be captured using deep learning.

    “A model using only aggregate step counts achieved an AUC of 0.7011 (diabetes), 0.5811 (sleep apnea), 0.5714 (hbp), and 0.5601 (high cholesterol), suggesting that although overall physical activity plays an important role in predicting the onset of disease, the heart's response to physical activity is a more salient biomarker that can be captured using deep learning.”
    Quote in the paperSubstantial claim

  • May combine two claims.

    Premisecitation
    § Introduction · p. 1

    Resting heart rate and heart rate variability correlate with diabetes, sleep apnea, atrial fibrillation, heart failure, sudden cardiac death, and irritable bowel syndrome.

    “including rich signals like resting heart rate and heart rate variability, which have been shown to correlate with health conditions as diverse as diabetes, sleep apnea, atrial fibrillation, heart failure, sudden cardiac death, and irritable bowel syndrome (Kamath, Watanabe, and Upton 2012).”
    Quote in the paperSubstantial claim0 of 1 sources checked
    • cites Kamath, Watanabe, and Upton 2012

      Couldn’t be matched to a paper in Crossref, so it can’t be checked.

  • Probably not a substantial claim: may not pass S2 (checkable). May combine two claims.

    Interpretationreasoning
    § Effect of Varying Amounts of Labeled Data · p. 5

    Wearables could be used to focus diagnostic testing on a high-risk subset of the population, leading to cost-effective screening for metabolic syndrome and getting at-risk patients care sooner.

    “This suggests wearables could be used to focus diagnostic testing on a high-risk subset of the population, leading to cost-effective screening for metabolic syndrome and getting at-risk patients care sooner.”
    Quote in the paperNot a substantial claim

  • Probably not a substantial claim: may not pass S2 (checkable). May combine two claims.

    Interpretationcitation
    § Discussion · p. 5

    Since these four conditions are highly prevalent, commonly undiagnosed, and reduced HRV is associated with the earliest stages of disease progression, this suggests a new approach to public health screening using deep learning.

    “Since these four conditions are highly prevalent, commonly undiagnosed, and reduced HRV is associated with the earliest stages of disease progression (Schroeder et al. 2005; 2003), this suggests a new approach to public health screening using deep learning.”
    Quote in the paperNot a substantial claim0 of 1 sources checked
    • cites Schroeder et al. 2005; 2003

      Couldn’t be matched to a paper in Crossref, so it can’t be checked.

  • Classified as probably not a substantial claim. May combine two claims.

    Premisecitation
    § Discussion · p. 6

    Statistical models in medicine are typically used for etiology, not just prediction, and cause and effect may be intertwined in subtle ways within medical data sets.

    “Statistical models in medicine are typically used for etiology, not just prediction, and cause (disease) and effect (treatment) may be intertwined in subtle ways within medical data sets (Shmueli and others 2010).”
    Quote in the paperSubstantial claim0 of 1 sources checked
    • cites Shmueli and others 2010

      Couldn’t be matched to a paper in Crossref, so it can’t be checked.

18 of 18

  • The paper relies on this without support.

    No such database exists for wearable data.

    “No such database exists for wearable data, and in most hospital systems, recruiting even a hundred patients into a new medical study is labor-intensive and expensive.”

  • The paper relies on this without support.

    Recruiting even a hundred patients into a new medical study is labor-intensive and expensive.

    “No such database exists for wearable data, and in most hospital systems, recruiting even a hundred patients into a new medical study is labor-intensive and expensive.”

  • The paper relies on this without support.

    High cholesterol has the least direct physiological relationship with heart rate variability in the medical literature.

    “Among all four health conditions, high cholesterol has the least direct physiological relationship with heart rate variability in the medical literature.”

  • The paper relies on this without support.

    Purely supervised deep learning requires many labeled examples.

    “However, purely supervised deep learning requires many labeled examples.”
4 of 4

Claims other papers make about this one

Each is checked against this paper’s own text: does it say what the citing paper says it does?

  • Recent advances in deep learning architectures for sequential modelling based upon wearable and mobile sensing have been used for health predictions and recommendations.

    “Recent advances in deep learning architectures for sequential modelling based upon wearable and mobile sensing have been used for health predictions and recommendations [141].”

    From Human-centred artificial intelligence for mobile health sensing: challenges and opportunities (2023) · cites this paper as [141]

Self-check

Percentages, sums and reported differences re-computed from the paper’s own numbers.

Each contradiction between two passages listed here was confirmed by an independent check.

  • Data-set split counts exceed the stated total

    Major

    The reported training, tuning, and validation counts sum to 64,973 person-weeks, not the stated total of 57,675.

    Reported 57675 · recomputed 64973

  • Conflicting results for heart-rate-only prediction of high blood pressure

    Minor

    The same passage says the heart-rate-only model performed significantly worse on high blood pressure and also performed the same within the 95% confidence interval.

2 of 2

Cites

Flagged first, then by citations

Cited by

Sorted by citations

Related

Authors
Ballinger, B., Hsieh, J., Singh, A., Sohoni, N., Wang, J., Tison, G., Marcus, G., Sanchez, J., Maguire, C., Olgin, J., & Pletcher, M.
Affiliations
  • Stanford Medicine (US)
  • University of California, San Francisco (US)
ORCID
Venue
Proceedings of the AAAI Conference on Artificial Intelligence
Publisher
Association for the Advancement of Artificial Intelligence (AAAI)
Published
2018-04-26
Volume
32
Issue
1
ISSN
  • 2159-5399 (ISSN-L)
  • 2374-3468
DOI
10.1609/aaai.v32i1.11891
OpenAlex
W2785925437
Cited by
142
References
40
Citations by year
2026: 8 · 2025: 8 · 2024: 7 · 2023: 22 · 2022: 17 · 2021: 24 · 2020: 24 · 2019: 24 · 2018: 8
FWCI
23.24
Type
conference paper
Language
en
Open access
Locations
  • Proceedings of the AAAI Conference on Artificial Intelligence (publishedVersion)
  • AAAI Publications (The Association for the Advancement of Artificial Intelligence (AAAI)) (submittedVersion)
Topics
  • Health Sciences › Health Professions › Health Information Management › Artificial Intelligence in Healthcare
  • Physical Sciences › Engineering › Biomedical Engineering › Non-Invasive Vital Sign Monitoring
  • Physical Sciences › Computer Science › Artificial Intelligence › Machine Learning in Healthcare
Keywords
cardiovascular risk prediction, semi-supervised learning, sequence learning, LSTM, wearable heart rate sensors, diabetes detection, hypercholesterolemia, hypertension, sleep apnea, multi-task learning, risk stratification, pre-training

About these checks

Registry, retraction, Cites and statistics results are computed, not judged, with one exception: a printed reference that no index lists is read once by a model before it counts as a citation error. Claims are extracted by a model and kept only when their quote appears verbatim in the text; each cited source is then checked, with a model’s help, for whether it says what the claim relies on. Those verdicts can be wrong — read the claim’s quote and the cited source before relying on one.

Claims read by deepseek-v4.1-flash under definition version 3 (extract-v3) on 2026-10-09.

Classified by jev-1.13-20260917 (gold-guideline-v0.2) on 2026-10-09.

Cited sources are checked by gpt-6-sol.

Statistics and arithmetic are recomputed, not judged.

Confidence bands are shown only for fields calibrated on our annotated test set. None is calibrated yet, so no band is shown.

What counts as a claim ›How we measure confidence ›How we check this ›

Are you an author, or do you see a problem with this page? Request a review