Skip to content
Wearables & measurementCohort study4 min read

Fitbit sleep heart rate flagged hypothyroidism with 74.6% accuracy in a mostly female cohort

MoveVitals Editorial Collective · Study published · Reading updated

Original paper:Harnessing AI for Hypothyroidism Monitoring Using Wearable-Derived Heart Rate Data: Retrospective Cohort Study. · JMIR AI · 2026 · Original paper ↗

Key takeaway

A machine-learning model using 15 nights of Fitbit sleep heart rate plus prior thyroid blood tests classified thyroid status with 74.6% accuracy (sensitivity 82%, specificity 73.5%) in the harder euthyroid-versus-hypothyroid subset (744 paired observations). The 413 participants were 76% women and largely in a hormone-withdrawal setting, so it is unproven in typical primary hypothyroidism.

Study design

Cohort study

Reading basis

Published abstract

Reading focus

Association & confounding

The research question

Hypothyroidism starts slowly with vague symptoms, so diagnosis is often delayed, and repeated blood tests are costly and invasive. Because thyroid hormone affects heart rate, the authors tested whether wearable sleep heart rate could help flag at-risk people who should get a confirmatory blood test.

Who and how

Retrospective cohort study with 1677 paired observations of sleep heart rate and blood thyroid tests (free thyroxine and thyroid-stimulating hormone) from 413 participants (mean age 43.1, SD 9.7 years; 314 women, 76%) who wore Fitbit devices. Each blood test was paired with sleep heart rate from the 15 consecutive nights before it. A gradient boosting model (LightGBM) combined baseline thyroid results with statistical and temporal heart rate features and was assessed with patient-level leave-one-out cross-validation. Performance was compared using 1, 2 or 3 earlier paired observations as reference points.

What it found

The authors’ reported results. Read each estimate with its population and uncertainty.

  1. 1
    In the main analysis of euthyroid and hypothyroid cases only (744 paired observations), the model reached 74.6% accuracy, 82% sensitivity and 73.5% specificity with one reference pair.
  2. 2
    Using 3 reference pairs improved accuracy to 78.1% (sensitivity 81.8%, specificity 77.7%), so longer personal history helped.
  3. 3
    In the full dataset (1677 paired observations), accuracy was 80.8% with 82% sensitivity and 80.8% specificity using one reference pair. This set includes hyperthyroid cases, which the authors note can inflate performance.
  4. 4
    The model detected 96.3% of severe overt hypothyroidism cases (free T4 <0.5 ng/dL).

Limits & uncertainty

  1. 1
    The authors describe the cohort as predominantly female and an iatrogenic-withdrawal group, meaning hypothyroidism was largely induced by stopping thyroid hormone, so it may not reflect people with naturally developing hypothyroidism.
  2. 2
    Leave-one-out cross-validation tests the model within one dataset; it was not checked in an independent group or prospectively.
  3. 3
    The model also uses baseline blood test results as input, so it tracks change from known values rather than screening from heart rate alone.
  4. 4
    At 74.6% accuracy in the harder subset, many people would still be misclassified; this reading is based on the abstract only.

What to take from it

Overnight heart rate from a wearable may carry a weak signal about thyroid status, and models like this might one day prompt people to get a blood test. It is not a way to diagnose or rule out a thyroid problem. Symptoms such as fatigue, feeling cold or an unusually slow heart rate are worth raising with a clinician, who can order the proper tests.

Reading your own data

Nightly or sleeping heart rate is recorded by most wearables. Track your own trend over weeks on the same device, worn the same way, and note it alongside symptoms and any thyroid results you have. A drift in sleeping heart rate has many possible causes, so use it as a prompt for a conversation with your clinician, not as a thyroid test.

How to read this evidence

Exposure precedes follow-up outcomes, but confounding and reverse causation can remain. A hazard ratio is not an individual absolute risk or a demonstrated intervention effect.

Authors’ structured abstract English

The attributed English abstract is reproduced under the recorded open licence. It is source material, not a MoveVitals interpretation. The prepared analysis above separates the authors’ findings from our interpretation.

Background

Hypothyroidism is a prevalent endocrine disorder characterized by an insidious onset and nonspecific symptoms, often leading to significant diagnostic delays. While biochemical thyroid function tests (TFTs) are the gold standard, their invasive nature and cost make them impractical for continuous, proactive screening of at-risk populations. Thyroid hormones exert direct chronotropic effects on the cardiovascular system, and their levels are closely correlated with heart rate (HR). Wearable devices capable of continuous HR monitoring offer a novel opportunity for noninvasive, digital phenotype-based detection.

Objective

This study aimed to develop and validate an AI-assisted system that uses wearable-derived sleep HR data as a digital monitoring tool for individuals at risk of hypothyroidism, helping identify those who warrant further confirmatory thyroid testing.

Methods

This retrospective cohort study used 1677 paired observations consisting of sleep HR data and serum TFT results, including free thyroxine (T4) and thyroid-stimulating hormone levels, from 413 participants (mean age 43.1, SD 9.7 y; women: n=314, 76%). Participants wore consumer-grade wearables (Fitbit) that recorded HR and sleep data. Each TFT result was paired with sleep HR data extracted from the 15 consecutive nights preceding the blood test. An AI system was developed using a gradient boosting algorithm (LightGBM) to classify thyroid status by integrating baseline TFT results with statistical and temporal HR features. The model's performance was evaluated using patient-level leave-one-out cross-validation. To examine the benefit of longitudinal data, performance was compared using 1, 2, or 3 prior paired observations as reference points.

Results

To ensure clinical relevance and avoid performance inflation from hyperthyroid cases, we prioritized the analysis of a subset containing only euthyroid and hypothyroid data (n=744 paired observations). In this clinically challenging subset, the system achieved an accuracy of 74.6%, with a sensitivity of 82% and a specificity of 73.5% using a single-reference data pair. Incorporating longitudinal HR patterns further enhanced performance; using 3 reference pairs improved accuracy to 78.1% (sensitivity 81.8% and specificity 77.7%). In the overall dataset (n=1677 paired observations), the system demonstrated an accuracy of 80.8%, with a sensitivity of 82% and a specificity of 80.8% using a single-reference pair. Notably, the system achieved high clinical safety by detecting 96.3% of severe overt hypothyroidism cases (free T4 <0.5 ng/dL).

Conclusions

This AI-assisted system demonstrates the feasibility of a sensitive, noninvasive longitudinal monitoring tool for hypothyroidism, robustly identifying a high proportion of hypothyroid cases-particularly severe ones-to prioritize at-risk individuals for confirmatory blood tests. Given the predominantly female iatrogenic-withdrawal cohort, prospective sex-balanced validation in primary hypothyroidism is warranted.

Sources & attribution

Harnessing AI for Hypothyroidism Monitoring Using Wearable-Derived Heart Rate Data: Retrospective Cohort Study.

Shin K, Jung Y, Yoo WS, Kim MJ, Oh TJ, Moon JH, Kong SH, Lee WW, Moon JH.

JMIR AI · 2026-10-28

DOI: 10.2196/86030 ↗PubMed · 42950109 ↗Europe PMC · Recorded licence: cc by ↗

Abstract © the original authors, reproduced under the recorded open licence. Translations support reading; use the English record to check terminology and numerical results. Our methodological interpretation is separate from the authors’ conclusions.

MoveVitals editorial interpretation · educational content · no individual clinical review recorded

Sources checked 2026-11-01

Practical guides

Read in context

·Cohort study

Activity–stroke and heart attack curves differ by accelerometer processing method

Among 90,237 UK Biobank adults free of cardiovascular disease at baseline, followed until December 2022 for 1,298 strokes and 2,031 myocardial infarctions, the shape of the link between moderate-vigorous physical activity (MVPA) and each outcome depended on how the accelerometer data were processed. For stroke, hazard fell in a straight line only with the machine-learning metric; for myocardial infarction, two metrics gave a curved decline and activity counts a straight one. The abstract gives no hazard ratios, so it shows that methods disagree, not which one is right, and the findings are observational associations.

Preventive Medicine

In-depth explainersRead study →
·Validation study/review

Smartwatch data as forensic evidence: heart rate is best validated, the rest uneven

A systematic review and meta-analysis of studies on smartwatches from four major brands, framed around digital forensics, found that heart rate was both the most studied and the most accurate measure. Step count, distance, energy expenditure, activity classification and sleep were less consistently validated, and results varied by brand, protocol and reporting. The abstract gives no pooled numbers, and most studies were run in healthy people in laboratories, so it cannot say how a particular watch performs for a particular person in daily life.

Forensic Science International

In-depth explainersRead study →
·Overview of reviews

Activity trackers: about 1,800 more steps a day, but small effects on other outcomes

An umbrella review of 39 systematic reviews and meta-analyses, covering 163,992 participants, found that activity trackers were associated with more physical activity (SMD 0.3–0.6), equating to about 1,800 extra steps and 40 minutes more walking per day, and with a bodyweight reduction of about 1 kg. Effects on blood pressure, cholesterol, glycosylated haemoglobin, quality of life and pain were typically small and often non-significant. The abstract does not say what else came with the trackers in these programs or how much the reviews overlap, so the gains cannot be credited to the device alone.

The Lancet Digital Health

In-depth explainersRead study →
·Validation study/review

Seven sleep trackers: good at spotting sleep, weak on wake and sleep stages

In a sleep laboratory, 34 healthy young adults wore a subset of seven consumer sleep trackers for three nights. Against polysomnography every device detected sleep very well (sensitivity 0.93 or higher) but caught wake poorly (specificity 0.18–0.54), and five of the seven matched or beat research actigraphy (specificity 0.39). Sleep-stage estimates were inconsistent, and the participants were healthy adults aged 18 to 35 in a fixed 8-hour bed window, with each device tested in only 11 to 20 of them, so the results say little about people with sleep problems.

Sleep

In-depth explainersRead study →
Wearables & measurement →