Sleep and temperature data from wearable devices support noninvasive detection of diabetes mellitus in a large-scale, retrospective analysis

Figures

• 

  Fig. 1. Population level comparisons between people self-reporting DM and people self-reporting no DM DX. a Number of people self-reporting DM, people self-reporting no DM DX, and No Response reports in our dataset. b Feature extraction of nightly features. The left panel shows one week of nightly measurements during sleep, with sleep intervals typically spanning midnight (time 0) and ending before noon (time 12). c–f Kernel density estimates of the commonly used features (hr_average, rmssd, total_sleep_time, diurnal_distal_body_temperature_amplitude) by each age group for both populations reporting DM DX status. “People self-reporting DM”: red; “People self-reporting no DM DX”: blue. Darker/more saturated lines indicate older age; lighter lines indicate younger age.

• 

  Fig. 2. Machine learning algorithms classify time windows of physiological data from people self-reporting DM. a Construction of machine learning datasets from cohorts of people self-reporting DM and people self-reporting no DM DX. Datasets vary in the length of input time windows. b–e ROC curves and PRC for each ML model used, over a range of different input time windows. Longer time windows are shown in darker shades. Gradient boosting classifier results are shown in purple. Logistic regression classifier results are shown in orange.

• 

  Fig. 3. Analysis of false positives and confounding variables in predictions from the best machine learning algorithm. a Confusion matrix for best performing model (21-night XGBClassifier). Darkness is proportional to the number of time windows in each category. b Average probability of a positive prediction in people self-reporting DM, people self-reporting no DM DX with other chronic conditions, and people self-reporting no chronic conditions. Red vs. Purple: p = 7.0 e-6; red vs. blue: p = 6.9 e-12 (see Supplementary Table 1). c Distributions of false negatives and true positives, shown in comparison to the true negative distribution. Exact p-values in Supplementary Table 2. All statistical test results shown in this figure were generated from Dunn’s Test, using Benjamini–Hochberg p-value threshold correction. * indicates a p-value < 0.05, ** indicates a p-value < 0.01, *** indicates a p-value < 0.001, and n.s. (not significant) indicates a p-value ≥ 0.05.

• 

  Fig. 4. Feature importance and ablation analyses of best machine learning algorithm. a Feature weights for each feature averaged across all the nights in the input window for the best performing model. b Mean improvement in classifier performance as a result of adding each feature set. Error bars represent 25th and 75th quantiles of the increase in AUROC upon addition of each set (n = 16 possible feature set additions per feature). c Reduction in classification errors in people self-reporting no DM DX with other chronic conditions, as a result of adding feature sets 4 and 5.

• 

  Fig. 5. Evaluation of the best model on an imbalanced cohort. ROC Curve (a) and PRC Curve (b) of best algorithm (21-night XGBClassifier) on the imbalanced cohort with a class balance of 17.9%. Results are the average of 100 repeated trials where individuals are randomly drawn from a pool of 6571 individuals who self-reported no DM DX. Shaded regions indicate the interquartile range of repeated experiments at each threshold.