Compared with the standard of care, the use of an algorithm that provides a personalized, data-driven treatment recommendation may significantly improve management of type 2 diabetes, according to a study published in Diabetes Care.
Researchers modeled treatment outcomes for 13 pharmacological therapies using information on historical outcomes from patients with type 2 diabetes and similar characteristics. This information was based on the electronic medical records (EMRs) of 10,806 patients from 1999 to 2014 at Boston Medical Center. The researchers then developed a nonparametric prescriptive treatment algorithm that offers recommendations tailored to an individual patient.
“For each patient visit, we used k-nearest neighbor (kNN) regression to predict the potential HbA1c outcome under each treatment alternative,” the researchers explained. “The nearest neighbors were chosen to control for confounding that may be present in nonrandomized data and to maximize similarity on the patient characteristics that were most predictive of outcomes. Then algorithm then prescribed the regimen with best predicted outcomes, provided the predicted improvement relative to the patient’s current regimen exceeded a confidence threshold.”
To evaluate the effect of the algorithm, the researchers compared expected HbA1c under the recommended therapy with observed HbA1c under standard of care.
For 68.2% of 48,140 patient visits in the dataset, the algorithm did not differ from standard of care. However, in the 31.8% of patient visits in which they did differ, mean HbA1c was 0.44±0.03% lower with the algorithm vs standard of care (P <.001), representing a decrease in HbA1c from 8.37% with standard of care to 7.93% with the algorithm.
Sensitivity analyses conducted under 3 alternate random splittings of the dataset showed that the overall mean benefit of using the algorithm vs standard of care ranged from 0.11% to 0.15% (P <.001 for all instances).
The researchers also discussed their prototyped dashboard for health care providers. The dashboard includes information on demographics, medical history, and response to treatment for patients similar to index patients, as well as visualizations of the patient’s treatment history and progression of HbA1c. The dashboard also displays the mean, standard deviation, and full distribution of HbA1c outcomes among the kt* nearest neighbors who received each treatment in the menu of options. The dashboard then offers a treatment recommendation that the health care provider can follow or override based on the patient’s circumstances.
“Compared with other machine-learning methods considered, the kNN prescriptive approach is highly interpretable and flexible in clinical applications,” the researchers wrote. “The novelty of our approach is in personalizing the decision-making process by incorporating patient-specific factors. This method can easily accommodate alternative disease-management approaches within specific subpopulations … We believe this personalization is the primary driver of benefit relative to standard of care.”
- Patients were not randomly assigned to treatment groups.
- EMRs do not include information on socioeconomic factors or patient preferences.
- Data were lacking on glucagon-like peptide-1 (GLP-1) agonists, which were therefore not deemed a separate drug class.
- Because they used EMR data only, the researchers could not be certain that patients filled and took their medications or precisely when a medication was stopped.
Disclosures: The researchers report no conflicts of interest.
- Bertsimas D, Kallus N, Weinstein AM, Zhuo YD. Personalized diabetes management using electronic medical records [published online December 5, 2016]. Diabetes Care. doi:10.2337/dc16-0826.