389-P: Ability for Detecting or Predicting Hypoglycemia with the Aid of Machine Learning Techniques: A Meta-analysis

Hypoglycemia is a major barrier to maintaining optimal glycemic control. Machine learning (ML) approaches have been applied to improve the identification of unrecognized hypoglycemia. This meta-analysis assessed current detective or predictive ability for hypoglycemia using ML techniques. Electronic literature searches were conducted using MEDLINE and EMBASE for detecting real-time hypoglycemia or predicting hypoglycemia subsequent to several features, where an index test is a model constructed by ML algorithms and the reference standard is a blood test to confirm the occurrence of hypoglycemia. Included studies were restricted to sample-based studies that examined both the ability to discriminate non-hypoglycemia events from hypoglycemia events and the ability to discriminate hypoglycemia events from non-hypoglycemia events. The 2x2 contingency data (i.e., true-positive, false-positive, false-negative, and true-negative) were pooled with a hierarchical summary receiver operating characteristic model. Pooled sensitivity/specificity (95% confidence interval (CI)) of 6 eligible studies for predicting hypoglycemia and 5 eligible studies for detecting hypoglycemia were 0.78 (0.70-0.85)/0.81 (0.69-0.89) and 0.78 (0.72-0.83)/0.84 (0.63-0.94), respectively. Consequently, the positive likelihood ratio (PLR) calculated as [sensitivity/(1-specificity)] and the negative likelihood ratio (NLR) calculated as [(1-sensitivity)/specificity] with 95% CI were 4.03 (2.26-7.18)/0.27 (0.17-0.42) and 4.90 (1.92-12.51)/0.27 (0.21-0.33), respectively. Considering that PLR>5 moderately increases the probability of a disease given a positive test and NLR<0.2 moderately decreases the probability of a disease given a negative test, it is concluded that current models have not yet attained satisfactory performance and that the algorithm may need further development before being implemented.