Wearable sensors enable personalized predictions of clinical laboratory measurements

Abstract

Vital signs, including heart rate and body temperature, are useful in detecting or monitoring medical conditions, but are typically measured in the clinic and require follow-up laboratory testing for more definitive diagnoses. Here we examined whether vital signs as measured by consumer wearable devices (that is, continuously monitored heart rate, body temperature, electrodermal activity and movement) can predict clinical laboratory test results using machine learning models, including random forest and Lasso models. Our results demonstrate that vital sign data collected from wearables give a more consistent and precise depiction of resting heart rate than do measurements taken in the clinic. Vital sign data collected from wearables can also predict several clinical laboratory measurements with lower prediction error than predictions made using clinically obtained vital sign measurements. The length of time over which vital signs are monitored and the proximity of the monitoring period to the date of prediction play a critical role in the performance of the machine learning models. These results demonstrate the value of commercial wearable devices for continuous and longitudinal assessment of physiological measurements that today can be measured only with clinical laboratory tests.

Extended Data Fig. 1 |. Wearables temperature variations and extended modeling results.

a, Variations in wRTemp over course of the day. b, R statistics based on LOOCV for all tests from Fig. 3b. c, R statistics based on K-fold CV for all tests from Fig. 3b.

Extended Data Fig. 2 |. Model accuracy changes over time based on window of historical data from an individual.

a, Lasso regularized regression using features calculated using different windows of wearable device monitoring. b, Accuracy of the HCT cVS mixed effects models over time for two example patients that were monitored between 2.5–5 years at Stanford hospital with >50 HCT observations at separate clinic visits. The HCT cVS mixed effects models demonstrate that the model accuracy changes over time, and particularly with a dramatic health event like a myocardial infarction (ICD code I21.4) (red vertical line) or a life-threatening ED visit (blue vertical line; CPT code 99285).

Extended Data Fig. 3 |. Increasing amounts of personalized data open up new study and model possibilities.

a, Summary of different biomedical data collection modalities and the typical amount of data they result in. b, Demonstration of how the amount and modality of data collection (longitudinal continuous vs. discrete measurements) constrain the type and complexity of models that can be built from the data.

Fig. 1 |. Overview of the iPOP wearables study.

a, Study design. b, Timespan of clinical monitoring per participant in the iPOP wearables cohort (left), and the total number of clinic visits per person (right). Each clinic visit included clinical lab tests. n = 54 study participants in each plot. c, Distribution of vital signs measured in the clinic and by the watch in the iPOP wearables cohort (n = 226 measurements). The values of wRHR and wRTemp were computed by averaging the wHR and wTemp during periods in which no steps were taken, including all such periods that occurred 2 weeks before clinic visits during the same time period as the clinic visits (7:00 to 9:00). Median values are indicated by dark blue vertical lines. d, Daily variation in median wRHR using multiple resting definitions (no steps or steps < 50 for a duration of 10 or 60 min) (n = 54 participants with at least one cHR and cTemp measurement (2,145 observations in total) during wearables monitoring). e, Variance of wRHR using multiple resting definitions (no steps for a duration of 60, 10 or 5 consecutive min). Measurements of wRHR are taken from hours of the day corresponding to typical clinic visit times for a duration of either 1 week, 2 weeks or 1 month before the clinic visit. The average variance of wRHR across the nine different resting definitions is 53.2 and the variance of cHR is shown as a horizontal line at 93.2 bpm. n = 54 participants.

Fig. 2 |. Methodology for predicting clinical laboratory measurements from vital signs collected using wearables.

a, Feature engineering pipeline to calculate potential digital biomarkers from continuous, longitudinal smart watch data. Statistical moments of the wVS, including heart rate, skin temperature, EDA and step counts, were subjected to thresholding based on the time of day, impact level of physical activity, and domain knowledge to reduce the size of the feature set. b, Overview of the modeling and analysis approach for this study, including the input data (left), statistical learning methods employed (middle) and model evaluation methodology (right).

Fig. 3 |. Predicting clinical laboratory measurements from vital signs collected using wearables.

a, Physiological categories of clinical laboratory tests performed at clinic visits. ALB, albumin; ALKP, alkaline phosphatase; ALRCU, aluminum/creatinine ratio; ALT, alanine aminotransferase; AST, aminotransferase; BASO, relative basophil count; BASOAB, absolute basophil count; BUN, blood urea nitrogen; CHOL, total cholesterol; CHOLHDL, high-density lipoprotein/total cholesterol ratio; CR, creatinine; EOS, relative eosinophil count; EOSAB, absolute eosinophil count; GLOB, globulin; HbA1c, glycated hemoglobin; HDL, high-density lipoprotein; HSCRP, high-sensitivity C-reactive protein; IGM, immunoglobulin M; LDL, low-density lipoprotein; LDLHDL, LDL/HDL ratio; LYM, relative lymphocyte count; LYMAB, absolute lymphocyte count; NEUT, relative neutrophil count; NEUTAB, absolute neutrophil count; MCH, mean corpuscular hemoglobin; MCHC, mean corpuscular hemoglobin concentration; MCV, mean corpuscular volume; NHDL, non-HDL cholesterol; RDW, red-cell distribution width; TBIL, total bilirubin; TGL, triglycerides; TP, total protein; UALB, urine albumin; WBC, white-blood-cell count. b, The models that most accurately predict clinical laboratory tests using vital signs measured by the watch (wVS, red triangles) compared to the clinic (cVS, blue and green circles) (P < 0.05 for all except serum chloride (CL); correlation between observed and predicted values with Bonferroni correction). Points correspond to the mean R statistic derived by leave-one-person-out cross validation for n = 54 study participants, and error bars represent the 95% confidence intervals derived by bootstrap with the procedure repeated 1,000 times. The wVS are random forest models using the 153 digital biomarkers from part c calculated on watch data from the day before the clinic visit. The cVS models are bivariate linear (blue) or random forest (green) models with cHR and cTemp as model features. All of the models are cross validated using leave-one-person-out cross validation and confidence intervals are established using bootstrapping (P < 0.05). Clinical laboratory test colors correspond to physiology subsets from part a. c, The most accurate digital biomarkers selected out of the 153 features in the wVS models in part b. The colors and large icon in the background of the squares correspond to the different wVS in the left side of Fig. 2a (pink heart, heart rate; blue droplet, EDA; tan thermometer, skin temperature; gray footprints, steps), and the foreground icons correspond to the thresholding criteria on the right side of Fig. 2a. Interpretations of colors and symbols are provided in part a. d, CCA using physiology categories from part a as outcome variables and the 153 digital biomarkers from Fig. 2a as model features (P < 0.05 for all CCA models). Points correspond to the mean correlation derived by leave-one-person-out cross validation for n = 54 study participants, and error bars represent 95% confidence intervals derived by bootstrap with the procedure repeated 1,000 times.

Fig. 4 |. Relationship between duration and proximity of monitoring and model accuracy.

a, The eight most accurate random forest models using varying time windows of wVS monitoring before the clinic test for calculating features as in Fig. 2a, and using leave-one-person-out cross validation for n = 54 study participants. Points correspond to the mean R statistic and error bars represent 95% confidence intervals derived by bootstrap with the procedure repeated 1,000 times. b, Multiple correlation coefficient (R) of the predicted versus observed values in the personal HCT cVS mixed effects model and wVS personal random forest model over time for the most frequently sampled iPOP study participant (a mainly healthy individual), with simultaneous smart watch monitoring and frequent clinic sampling over a 2.5-year period. The clinic visits demarcated with arrows correspond to a viral infection (left and middle arrows) and a traumatic biking accident resulting in an ED visit (right arrow). c, Accuracy (R) of the HCT ~ All Vitals model in the iPOP participant from part a versus the number of clinic visits that were used to develop the model.

Fig. 5 |. Personalized models improve predictions of clinical laboratory tests from vital sign measurements.

a, Comparison of five models predicting clinical laboratory test values in the SEHR dataset for patients with more than 50 observations for each clinical laboratory test (average n = 117 patients per test; the number of patients varies for each test). The models include the personal mean of the test for a patient (red), the linear clinic vitals (cVS) model (~All Vitals) (olive green), the personal mean + linear cVS model (green), the personal cVS random forest model (blue), and the linear mixed effects models using the personal mean and slope + cVS (purple). Points correspond to the mean R statistic derived by cross validation and error bars represent 95% confidence intervals derived by bootstrap, repeating the procedure 1,000 times. b, Study summary and results. Font sizes of clinical labs correspond to the overall predictive ability of the models developed in this study.