Counts describe this library, not the strength of evidence.
Comprendre ce type de preuve
Validation is specific to a metric, model, reference and setting. A useful heart-rate measurement does not validate calorie estimates or sleep stages. Even processing the same accelerometer data differently can change estimated activity–disease relationships.
In an online survey of 418 people with diabetes, 83.9% saw wearables as useful for monitoring glucose and activity, but 79.7% worried about data privacy, 77.0% about cost and 75.1% about usability; 81.9% would adopt AI-integrated devices if their healthcare provider recommended them. These are self-reported attitudes from a self-selected online sample, not evidence that the devices improve diabetes control.
Across 23 randomized trials with 4566 community-dwelling older adults, tracker-based interventions beat usual care for daily steps (SMD 0.58, 95% CI 0.33-0.83) and physical activity time (SMD 0.28, 95% CI 0.10-0.47) right after the intervention. BMI, body fat and physical function did not improve significantly, and the evidence was rated low to moderate certainty.
In an online survey of 455 people, most of them with chronic conditions, perceptions of AI-enabled wearable health devices were largely positive, but technical failures, data accuracy and less human contact were significant concerns. The sample came from convenience and stratified sampling and the abstract gives no figures, so how common each view is remains unknown.
In a 2022 survey of 5591 US adults, 36.36% (2033) used a wearable, and 78.4% of users (1584 of 2020) were open to sharing data with providers but only 26.5% (535 of 2020) had done so. This is a self-reported national survey at one point in time, so it describes patterns, not what drives them.
In this systematic review, studies pairing AI with wearable inertial measurement units (IMUs) had a median dataset of only 50 participants; 76% used traditional machine learning, 93% used supervised learning and 65% focused on neurological problems. The abstract does not give the number of studies reviewed or how well the models performed, so it describes research habits, not how well the tools work.
In 122 clinic patients, a handheld ECG, the Apple Watch ECG and two phone camera apps each detected atrial fibrillation (AF) with 100% sensitivity, and specificity for sinus rhythm ranged from 96.4% to 98.9%. Only 30 patients had AF, and 7.4% to 14.8% of recordings were of insufficient quality.
A systematic review of 23 papers from 12 distinct studies found some evidence that lower activity, higher heart rates and later, irregular sleep onset were linked to psychiatric diagnoses or poorer symptoms in schizophrenia or bipolar disorder. Devices, measures and statistics varied so much that results are hard to compare.
Across 23 pulse oximetry studies (n = 59,684) and 4 wearable pulse rate studies (n = 176), accuracy fell short of regulatory thresholds for light, medium and dark skin alike, with SpO2 mean bias highest in dark pigmentation (1.27%, 95% CI 0.58 to 1.95). Few wearable studies are included, and wearable pulse rate errors were widest for dark skin.
Across 38 studies, wearable AI detected sleep apnea with pooled accuracy 0.869, sensitivity 0.938 and specificity 0.752, but identified severity level with accuracy of only 0.651. The authors conclude performance is suboptimal for routine clinical use, and results varied with algorithm, data type, apnea type and device placement.
Among 2976 Shanghai schoolchildren without myopia at the start, only outdoor spells of at least 15 minutes at 2000 lux or brighter were associated with less myopic shift over 1 year (for example -0.007 D, 95% CI -0.011 to -0.002 D at 2000 to 3999 lux). The effects are tiny per exposure unit, and this is an observational analysis within a trial, so it does not prove that these patterns prevent myopia.
Across 24 systematic reviews covering 249 validation studies and 430,465 participants, consumer wearables had a mean heart rate bias of ±3% but overestimated VO2max by ±15.24% (resting tests) and ±9.83% (exercise tests). Only about 11% of released devices have been validated for any outcome, and the research is so varied in methods that firm accuracy verdicts are not yet possible.
In a US Census-matched survey of 23,974 people, 10,679 (44.5%) owned a wearable, and ownership was much lower among those aged 65 and older (OR 0.18, 95% CI 0.16-0.21) and the uninsured (OR 0.41). It is a self-reported consumer survey, so it describes who owns wearables rather than who benefits from them.
A scoping review of 10 studies of health care professionals found no physiological measure reliably associated with burnout or anxiety, though step count and time in bed were linked to depressive symptoms, and heart rate and heart rate variability to acute stress. The studies were few, short and small, so wearables for burnout prediction remain unproven.
In 30 people after a recent heart attack, a 26-week wearable-monitored online rehab program improved daily steps and daily distance more than gym-based rehab at 24 weeks (both p < 0.05), and both groups improved on all measures (p < 0.001). With only 30 participants, the claim that the two are equally effective is tentative.
Using Fitbit data from All of Us, individualized random forest models separated prepregnancy, pregnancy, postpartum without depression and postpartum depression periods (mAUC = 0.85; κ = 0.80). The cohort with postpartum depression (PPD) had fewer than 20 women, so the results are preliminary.
Among 1074 US adults over 65 with accelerometer data, greater day-to-day variability in sleep efficiency was linked to lower scores on the Digit Symbol Substitution Test (adjusted β = -2.01 per 10% increase) and Animal Fluency Test (-0.84), but not the word-learning test. The data are cross-sectional, so poor cognition could also disturb sleep.
In 8416 Swedish adults aged 50-65, time spent sedentary and time in moderate or vigorous activity were linked to the same set of gut bacterial species, but in opposite directions. Because the data are cross-sectional, the study cannot say whether activity changes the microbiome or whether people with different microbiomes move differently.
In a randomized trial of 220 recreational runners, real-time insole feedback did not significantly change injury rate in the intention-to-treat analysis (HR 1.11, P = .70). Because 53 of 160 analyzed participants (33%) used incorrect settings, the lower rate in the as-treated analysis (HR 0.53, P = .03) is less reliable.