JMIR Publications

ABSTRACT

Background:

Machine learning (ML) applications have been widely introduced to diabetes research, in particular, identifying hypoglycemia.

Objective:

This meta-analysis aimed to assess the current ability of ML algorithms to detect or predict hypoglycemia.

Methods:

We systematically searched cohort studies published from 1950 Jan.1 to 2019 Nov.18 using MEDLINE and EMBASE. Included studies had to compare the consistency between ML’s classification and diagnosis of hypoglycemia using blood tests. The set of 2x2 data (i.e., number of true-positives, false-positives, true-negatives, and false-negatives) in each study was pooled with a hierarchical summary receiver operating characteristic model.

Results:

Twenty studies (9 studies for detecting hypoglycemia (i.e., no interval between the ML’s classification and blood glucose test) and 11 studies for predicting hypoglycemia (i.e., data that were based on ML’s classification were examined before the blood glucose test) met the inclusion criteria. For detection of hypoglycemia, pooled estimates (95% confidence interval (CI)) of sensitivity, specificity, positive likelihood ratio (PLR) and negative likelihood ratio (NLR) were 0.76 (0.72-0.80), 0.66 (0.51-0.78), 2.20 (1.46-3.32), and 0.37 (0.28-0.49), respectively. For prediction of hypoglycemia, the pooled estimates (95% CI) were 0.84 (0.77-0.89) for sensitivity, 0.90 (0.84-0.93) for specificity, 8.05 (4.79-13.51) for PLR, and 0.18 (0.12-0.27) for NLR.

Conclusions:

Current ML algorithms have insufficient ability to detect hypoglycemia and moderate ability to predict hypoglycemia, judging from PLR and NLR values. Continued research is required to develop more accurate ML algorithms than currently exist and enhance the feasibility of applying ML in clinical settings. Clinical Trial: the international prospective register of systematic reviews (PROSPERO) (ID: CRD42020163682).