Objective
The overall objective of this project is to validate and, if possible, improve the performance of the IQVIA Adult T1D detection algorithm using EHR and genetic data from Geisinger health system. The specific aims are as follows:
1. Build and package production grade code for deploying the already developed Adult T1D Detection algorithm in Geisinger environment.
2. Apply the IQVIA Adult T1D algorithm to retrospective Geisinger EHR data and determine the classification performance; when available, validate using adult T1D cases confirmed by biomarker data.
a. Relevant patient counts from Geisinger data: GAD65 antibody= 4848, IA2 antibody=143, ZNT8 antibody= 30, c-peptide= 1695.
b. Through the funded JDRF project “Using genetics to better diagnose and treat adult-onset T1D in multi-ethnic US population” GADA, IA2A and Zn8TA antibody and c-peptide assays will be performed on 10,000 Geisinger patients who were treated with insulin, including those with clinically-suspected T1D and clinically-suspected T2D. These data should be available by the end of CY 2023.
3. Train the IQVIA algorithm using Geisinger retrospective EHR data to improve, if possible, the algorithm performance; identify modifications that improve the performance; collaborate with 3 additional sites performing algorithm validation to identify site-specific variables.
4. Develop a hybrid adult T1D risk model that combines the IQVIA EHR-based algorithm (best model from aims 1 and 2) and the genetic risk model; compare the performance of EHR based, genetic, and hybrid risk models.
a. Relevant patient counts from Geisinger data:
i. Genetic risk model: ~175k patients currently (potentially upwards of 200k patients by the time of the study)
ii. EHR-based model: >2M patients
5. For the best-performing algorithm from Aim 4 perform a prospective validation at Geisinger; develop a plan to implement the model in the clinical workflow; collaborate with other sites validating the IQVIA algorithm to assess generalizability.
Background Rationale
JDRF and IQVIA have developed the Adult T1D Detection Algorithm, a tool that is driven by AI analytics trained to identify patients who would benefit from additional diagnostic testing to confirm their T1D diagnosis.
This research project will build upon prior work to explore whether (1) further improvements can be made to Adult T1D Detection algorithm to boost precision by including additional EHR or genetic features and (2) significant customization of the deployment approach (both technical and clinical) is required for deployment at scale. This second goal will be achieved by combining the results of this research with a parallel workstream funded by JDRF to test the deployment of Adult T1D Detection algorithm at three other US health systems. Ultimately these projects are all being conducted with the goal of improving patient care by speeding a T1D diagnosis and providing patients with the appropriate care on an accelerated timeline.
Type 1 Diabetes (T1D) was once thought to be a childhood onset disease with a strong familial pattern, but it is now estimated that 90% of T1D patients do not have a family history of the disease, and almost half of newly diagnosed cases occur in adults. Nearly 40% of adults with T1D are initially misdiagnosed with T2D, leading to ineffective care (ex. high risk of diabetic ketoacidosis). By shortening the time to an accurate T1D diagnosis, patients will be able to have accelerated access to the correct treatments for their condition.
Description of Project
This proposal would provide funding for a collaboration between JDRF, IQVIA, and Geisinger Health System. Through separate initiatives with these organizations, JDRF has sponsored the development of predictive algorithms that detect adult patients who are on the journey to a final T1D diagnosis but may currently be diagnosed with T2D. Ongoing IQVIA work has resulted in a detection algorithm that leverages ambulatory EHR data. JDRF funded Geisinger Health System/Exeter projects have resulted in an algorithm that leverages genetic data and, separately, establishes gold standard T1D diagnoses that can be used to measure the efficacy of predictive tools.
This body of work will compare and contrast the effectiveness of those algorithms when fitted to retrospective Geisinger Health System patients and then proceed to develop a hybrid model that leverages both electronic health record data and genetic data. In the project's final phase, the best performing algorithm will be deployed into real patient care and then evaluated for its effectiveness over a period of several months.
Anticipated Outcome
The expected outcome of this opportunity is that IQVIA, JDRF, and Geisinger Health System develop and deploy a hybrid predictive algorithm into live patient care environments at Geisinger Health System which identifies, with a high degree of accuracy, patients that will be revealed to have T1D upon further diagnostic testing. This hybrid predictive algorithm is likely to have a higher degree of accuracy than either the genetic or the EHR-based models alone. Furthermore, in comparing this project to other ongoing IQVIA/JDRF work, insights will be gained as to the generalizability of different predictive algorithms which will inform any future algorithm deployments to a broader healthcare organization audience.
Relevance to T1D
Type 1 Diabetes (T1D) was once thought to be a childhood onset disease with a strong familial pattern, but it is now estimated that 90% of T1D patients do not have a family history of the disease, and almost half of newly diagnosed cases occur in adults. Nearly 40% of adults with T1D are initially misdiagnosed with T2D, leading to ineffective care (ex. high risk of diabetic ketoacidosis). By shortening the time to an accurate T1D diagnosis, patients will be able to have accelerated access to the correct treatments for their condition.