1051-P: Improving Glycemic Outcomes in T1D Using an Automated Decision Support Recommender System: Evaluation In Silico and Compared with Physician Recommendations

Most people with type 1 diabetes use multiple daily injections (MDI) therapy, yet there are limited decision support tools for this population. We designed and evaluated a K-nearest-neighbors decision support system (KNN-DSS) that utilizes continuous glucose data and Bluetooth-enabled insulin dose capture devices to detect problematic glycemic patterns and recommend insulin dosage adjustments to maximize time-in-range (70-180 mg/dL) and minimize hypoglycemia (< 70 mg/dL). The KNN-DSS employs distance and class weighting, a carb ratio aggressiveness algorithm, and an expert-opinion quality-control algorithm to filter erroneous recommendations.

The KNN-DSS was evaluated both in silico and in comparison with physician recommendations. In silico virtual patients (n=30) of varying weight, total daily insulin requirement, insulin sensitivity, and compliance with care were evaluated in a 1-year study whereby they received weekly recommendations. The primary outcome measures were the % time-in-range and % time-in-hypoglycemia. The KNN-DSS was further evaluated for accuracy compared with endocrinologist recommendations collected during a 4-week, at-home clinical trial of CGM-augmented MDI therapy (N = 12, 7 female, 30.8 ± 5 years, 79.6 ± 18.9 kg), and compared using Dice similarity coefficient.

In silico, the KNN-DSS improved time-in-range from 66.8% at baseline to 81.0% (21% relative improvement, p <0.005), and reduced time-in-hypoglycemia from 3.14% at baseline to 1.37% (56% relative improvement, p < 0.005) after one year. Compared to endocrinologists, the KNN-DSS recommendations to real-world patients were within acceptable agreement 75% of the time, 6% contraindicated, and 19% whereby KNN-DSS gave a recommendation when the physician gave no recommendation. KNN-DSS improves glycemic outcomes in silico and exhibits acceptable agreement with physician recommendations. Future clinical studies are planned.