Ensemble machine learning methods with data from wrist-worn wearable devices provide the best prediction of changes in glycemic control in adults with prediabetes, according to a study in NPJ Digital Medicine.

The randomized trial investigated the use of data from waist-worn wearable devices compared with wrist-worn wearables to improve risk prediction models for changes in glycemic control in adults with prediabetes during a 6-month remote-monitoring period.

The study authors obtained information on participants’ demographics, medical history, and laboratory testing. The information was used to fit prediction models to investigate machine learning methods vs traditional regression models; baseline information with wearable data vs baseline information alone; and data from wrist-worn wearables vs waist-worn wearables. The main outcome measure was the change in glycated hemoglobin (HbA1C).

A total of 186 study participants were included—93 used a waist-worn device and 93 used a wrist-worn device. Overall, the participants had a mean (SD) age of 56.7 years, 66.1% were women, and 69.4% were White. They had a mean body mass index of 32.7 (7.3) kg/m2 and mean baseline HbA1C of 6.1 (0.2).

In the waist-worn wearable group, 74 of 93 (77.7%) participants had end-of-study laboratory testing. Of this group, 5 (6.8%) had an increase in their HbA1C level of ≥0.3, and 14 (18.9%) had a decrease of ≥0.3. In the wrist-worn wearable group, 73 of 93 (78.5%) of the participants had end-of-study laboratory testing. Of these, 11 (15.1%) had an increase in HBA1C levels of ≥0.3, and 12 (16.4%) had a decrease of ≥0.3.

In the continuous model using the standard approach with only baseline information and traditional linear regression, no difference was observed in prediction. In the enhanced approach with wearable data added to traditional linear regression, the study participants in the wrist-worn arm showed significantly greater prediction (R squared in waist-worn arm, 0.41; 95% CI, 0.395-0.420; R squared in wrist-worn arm, 0.50; 95% CI, 0.491-0.515; P < .001). Prediction improved when ensemble machine learning was used in the wrist-worn arm (R squared in waist-worn arm, 0.66; 95% CI, 0.658-0.671; R squared in wrist-worn arm, 0.70; 95% CI, 0.694-0.714; P < .001).

In the binary model that assessed worsening glycemic control, the standard approach with baseline information and traditional logit regression found no difference in prediction between the wrist- and waist-worn arms. arms. In the enhanced approach with wearable data added to the traditional logit regression, the wrist-worn arm had significantly greater prediction (area under the curve [AUC] in the waist-worn arm, 0.55; 95% CI, 0.48-0.61; AUC in the wrist-worn arm, 0.74; 95% CI, 0.68-0.79; P < .001). Prediction improved when ensemble machine learning was used in the wrist-worn arm (AUC in the waist-worn arm, 0.68, 95% CI, 0.61-0.74; AUC in the wrist-worn arm, 0.85; 95% CI, 0.79-0.90; P < .001).

Similar results were observed in the binary model that assessed improvement of glycemic control with the greatest prediction in the enhanced approach with wearable data that used ensemble prediction (AUC in waist-worn arm, 0.72; 95% CI, 0.66-0.77; AUC in wrist-worn arm, 0.84; 95% CI, 0.77-0.91; P = .01).

The investigators noted that their study is based on a small sample of participants from a single health system. They also noted the study was conducted in a 6-month period; longer studies are needed to determine if their observations only reflected short-term changes in HbA1C. Additionally, information on diet was not included in the study.

“We found that consistently in all 3 models, prediction improved when (a) machine learning was used vs traditional regression, with ensemble methods performing the best; (b) baseline information with wearable data was used vs baseline information alone; and (c) wrist-worn wearables were used vs waist-worn wearables,” the investigators concluded.

Disclosure: Some of the study authors declared affiliations with technical device and other private companies. Please see the original reference for a full list of authors’ disclosures.

Reference

Patel MS, Polsky D, Small DS, et al. Predicting changes in glycemic control among adults with prediabetes from activity patterns collected by wearable devices. NPJ Digit Med. 2021;4(1):172. doi:10.1038/s41746-021-00541-1