Machine learning (ML) methods receive increasing attention for diabetes risk assessment. This meta-analysis quantitatively assessed the predictive ability of ML algorithms for incident T2DM. Electronic literature searches were done using MEDLINE and EMBASE for longitudinal studies whereby an index test was a model constructed by ML algorithms for predicting incident T2DM. Reference standard was a blood test or physicians’ diagnosis confirming T2DM. The model was restricted to algorithms that required the aid of ML (e.g., neural network, support vector machine, etc.). 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 and specificity (95% confidence interval (CI)) of 8 eligible studies were 0.70 (0.66-0.74) and 0.78 (0.69-0.85), respectively. (Figure) Positive likelihood ratio (PLR) calculated as sensitivity/(1-specificity) and negative likelihood ratio (NLR) calculated as (1-sensitivity)/specificity were 3.25 (95%CI, 2.25-4.58) and 0.38 (95% CI, 0.34-0.43), respectively. We conclude that existing prediction models are not yet satisfactory 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.
S. Kodama: None. T. Sato: None. M. Yamamoto: None. H. Ishiguro: None. M. Iwanaga: None. K. Fujihara: None. T. Yamada: None. K. Kato: None. H. Sone: Research Support; Self; Kyowa Hakko Kirin Co., Ltd., Novartis AG, Ono Pharmaceutical Co., Ltd., Taisho Pharmaceutical Co., Ltd., Takeda Pharmaceutical Co.