Inter- and intra-physician variability in insulin injection adjustments compared with Bayesian algorithm...

There is an unmet need for automated insulin adjustments for multiple daily injections in type 1 diabetes under unsupervised use. We aimed to assess the McGill decision support system (DSS), which comprises a Bayesian algorithm, by comparing its automated insulin adjustments with those made by endocrinologists.

We surveyed 13 Canadian endocrinologists who made mock insulin adjustments in three separate parts on retrospective participants' data. Part A (weekly data) and part C (biweekly data) compared the recommendations from physicians with those of the algorithm. Part B evaluated intra-physician variability by comparing the recommendations made by the same physician over time based on an identical dataset to part A.

In part A, the agreement rate (mean [SD]) on the direction of weekly adjustments between the algorithm and physicians was non-inferior to the agreement rate between physicians for prandial bolus (55% [5] vs 56% [7], respectively; p=0.006) and basal insulin (48% [6] vs 51% [11], respectively; p=0.037). Low full disagreement rates on weekly adjustments were also comparable between the pairs for prandial bolus (3.9% [2.3] vs 4.1% [2.3], respectively; p=0.006) and basal insulin (10.6% [4.6] vs 9.2% [7.1], respectively; p=0.23). Similar rates were observed for biweekly insulin adjustments in part C. When comparing intra-physician decisions in part A with those in part B, on average, physicians fully agreed with themselves 66% (SD 7) and 67% (SD 7) of the time for prandial bolus and basal insulin adjustments, respectively.

The direction of insulin adjustments was comparable between physicians and the McGill algorithm. The large intra-physician variability further emphasises the subjective nature of insulin management.