JMIR Publications

ABSTRACT

Background:

Lower extremity complications related to diabetic polyneuropathy (DPN) are under- detected and on the rise in the U.S. Clinical decision tools are needed to identify people at high risk for these complications.

Objective:

To determine if a machine learning enabled algorithm could adequately predict which patients were at highest risk for lower extremity complications within 24 months of an initial neuropathy diagnosis.

Methods:

The study was conducted across two large multispecialty healthcare systems in northern California and Colorado. We included 48,209 insured adults with a new DPN diagnosis (04/01/2012-12/31/2016) who were not near end-of-life in the 24 months before diagnosis. The outcome was assessed using a composite indicator for any diagnoses consistent with lower extremity complications (foot ulceration, osteomyelitis, gangrene, lower extremity amputation) occurring within 24 months of DPN diagnosis.

Results:

Lower extremity complications, primarily foot ulcerations, occurred in 4.8% of individuals. The SL risk estimator exhibited good discrimination (area under the Receiver Operating Characteristic curve (ROC): 0.845 [0.826, 0.863]) and calibration and outperformed a naïve logistic regression model using complete cases only (area under ROC: 0.804 [0.783, 0.825]). A modified SL algorithm, simplified to facilitate real-world adoption, had only slightly reduced discrimination (area under ROC: 0.817 [0.797, 0.837]) and outperformed the naïve logistic regression model in terms of precision gained relative to the frequency of alerts and number of patients needed to evaluate.

Conclusions:

A risk estimator for DPN-related lower extremity events exhibited good predictive accuracy and strong potential for clinical application. Additional research is needed to identify optimal contexts for deployment and strategies for maximizing algorithmic fairness.