A Novel Electronic Medical Record Search Method to Identify Patients With Ketosis-Prone Diabetes: Implications for Discovery of Atypical Diabetes

Abstract

We developed Python-based Expeditious Program for Parsing Electronic Records (PEPPER) as a novel electronic medical record (EMR) search tool. We tested its utility and efficiency to automate the first step of identifying patients with A⁻β⁺ ketosis-prone diabetes (KPD). Electronic charts of 1,660 youth with type 2 diabetes (T2D) were analyzed by PEPPER to identify those with diabetic ketoacidosis (DKA) within 6 months of diagnosis. The efficiency and accuracy of PEPPER were compared with manual review. Further review confirmed A⁻β⁺ KPD per the Rare and Atypical Diabetes Network criteria. PEPPER identified 110 youth with T2D and DKA, of whom 21 met full A⁻β⁺ KPD criteria. PEPPER significantly reduced chart review time for this initial critical step compared with manual searching (mean ± SD 13.4 ± 3.9 s vs. 26.6 ± 9.4 s per chart; P < 0.001), and was 100% accurate. PEPPER streamlines EMR review, significantly reducing manual effort without sacrificing accuracy.

Graphical Abstract

Introduction

Extracting data from electronic medical records (EMRs) is essential for clinical research, but existing tools in commercial EMR programs such as Epic largely depend on diagnostic codes such as those in the ICD9 or ICD10, which are insufficient to identify novel or emerging conditions. We describe here a new method, the Python-based Expeditious Program for Parsing Electronic Records (PEPPER), which can go beyond code-based identification of disease phenotypes, reduce the burden of manual review, and accelerate EMR-based searches.

As a proof of concept, we deployed PEPPER to automate the first and most labor-intensive step of chart review (1) to identify patients with an atypical diabetes phenotype, autoantibody-negative, β-cell function–positive ketosis-prone diabetes (A⁻β⁺ KPD), for enrollment into the Rare and Atypical Diabetes Network (RADIANT) (2,3). Because no standard diagnostic code yet exists for this novel condition, affected patients are usually coded as having type 2 diabetes (T2D).

Here, we attempted to identify youth with A⁻β⁺ KPD in the Epic EMR of Texas Children’s Hospital (TCH), Houston, among individuals with a diagnosis of T2D who had a clinical visit during the 3-year period from 2019 to 2022. Our primary objective was to evaluate the time efficiency of PEPPER in completing the initial screening step of identifying youth with diabetic ketoacidosis (DKA) within 6 months of diagnosis compared with manual review. As secondary objectives, we compared PEPPER’s accuracy with manual review in subset of the charts; furthermore, we characterized youth with and without DKA within 6 months of diagnosis and summarized the characteristics of patients meeting full RADIANT A⁻β⁺ KPD criteria.

Research Design and Methods

Participants

The study population included 1,660 youth under 23 years of age with a diagnosis of T2D documented in the TCH Epic diabetes flowsheet between August 2019 and August 2022. The diabetes flowsheet is a mandatory medical documentation tool completed by treating care providers after each outpatient diabetes visit and includes documentation of diabetes type and date of diabetes diagnosis. After extracting the list of youth with T2D from Epic along with the provider-documented diabetes diagnosis date, our aim was to identify patients meeting A⁻β⁺ KPD criteria using the RADIANT definition (Fig. 1). In our prior work (1), we found that criterion I (Fig. 1), identifying DKA at diagnosis or within 6 months, was the most time-consuming step. Therefore, we used PEPPER to automate this initial screen and reduce the number of charts requiring subsequent manual review. This study was approved and the requirement for informed consent waived by the Baylor College of Medicine Institutional Review Board (Houston, TX).

Figure 1.

RADIANT criteria for A⁻β⁺ ketosis-prone diabetes.

Algorithm Development

PEPPER was developed in Python. Details regarding algorithm development, error handling, and patient privacy protections are provided in Supplementary Materials.

Measurement of Time to Review Charts

A random sample of 100 of the 1,660 charts was used to compare the time required for manual review versus PEPPER to determine criterion I. PEPPER was then fine-tuned for screen coordinates, button locations, and loading delays before being applied to the remaining charts. A second reviewer completed the manual review, and the time was measured from opening the patient lookup window to entering the extracted result into the database and closing the chart.

Using PEPPER to Screen for KPD in Participants

Using a list of patients’ medical record numbers, PEPPER accessed each chart and navigated to admission histories. It screened for DKA hospitalization occurring at diagnosis or within 6 months thereafter and recorded the earliest such admission date, if present. Patients meeting criterion I then underwent manual review for the remaining RADIANT A⁻β⁺ KPD criteria.

Data Collection

PEPPER collected data on occurrence of DKA within 6 months of diagnosis and dates of DKA admission. The following demographic and clinical data were manually collected from patients’ medical records: age at diagnosis, sex, race or ethnicity, BMI percentiles, date of T2D diagnosis, autoantibody status (for insulin autoantibodies, glutamic acid decarboxylase-65 autoantibodies, islet cell antigen 512 autoantibodies, and zinc transporter-8 autoantibodies), C-peptide level, insulin dose, dates of last insulin administration, hemoglobin A_1c (HbA_1c) level, and occurrence of precipitating or provoking factors for DKA (e.g., pancreatitis, acute infections [4]) at the time of the DKA episode.

Statistical Analysis

Chart review times for PEPPER and manual review were compared for the same 100 charts. We compared baseline characteristics between patients with and without DKA within 6 months of T2D diagnosis. Continuous variables were compared using a Student t test or Mann-Whitney U test. Categorical variables were compared using the Pearson χ² test. Statistical analyses were performed using SciPy (5).

Data and Resource Availability

The data collected for this study can be made available to others upon request.

Results

Among 1,660 youth with T2D, the median age at diagnosis was 17 (interquartile range 15–20) years; 59.0% were female; 59.5% were Hispanic, 30.1% African American, 7.0% non-Hispanic White, 2.5% Asian, 0.5% Native American/Pacific Islander, and 0.4% were biracial. Overall, 6.6% (n = 110) had DKA at diagnosis or within 6 months. This yield was similar to that reported by Kubota-Mishra et al. (1) using manual screening methods (present study: n = 110 of 1,660 [6.6%] vs. Kubota-Mishra et al.: 56 of 770 [7.3%]; P = 0.557). Youth with DKA were more likely to be male and more likely to be African American or biracial (P < 0.001) (Table 1).

Table 1.

Characteristics of the study population based on DKA presentation within 6 months of diagnosis

Characteristic	Participants, n	Entire cohort (N = 1,660)	DKA at or within 6 months of diagnosis (n = 110)	No DKA at or within 6 months of diagnosis (n = 1,550)	P value (DKA vs. no DKA)
Age, median (Q1-Q3), years		17 (15–20)	17 (16–19)	17 (15–20)	0.43
Sex, n (%)	1,660				<0.01
Female		979 (59)	25 (32)	954 (60)
Male		681 (41)	52 (68)	629 (40)
Race or ethnicity, n (%)	1,629		76	1,553	<0.01
Non-Hispanic White		114 (7.0)	1 (1.3)	113 (7.3)
Hispanic		969 (59.5)	34 (44.7)	935 (60.2)
African American/Black		491 (30.1)	38 (50.0)	453 (29.2)
Asian		40 (2.5)	1 (1.3)	39 (2.5)
Native American/Pacific Islander		8 (0.5)	0 (0)	8 (0.5)
Biracial		7 (0.4)	2 (2.6)	5 (0.3)
BMI percentile, median (Q1-Q3)	919	98.9 (96.71–99.35)	98.64 (97.3–99.5)	98.9 (97.2–99.5)	0.71
Positive antibody status, n (%)	1,654	154 (9.3)	43 (39.1)	147 (9.3)	0.93
Age at diagnosis, median (Q1-Q3), years	1,654	13 (11–14)	13 (12–15)	13 (11–14)

PEPPER significantly reduced the time required to review charts for criterion I compared with manual review (P < 0.001). Manual review required 26.6 ± 9.38 (mean ± SD) s per chart, translating to a projected 12.27 h to review all 1,660 charts. PEPPER required 13.41 ± 3.93 s per chart, corresponding to 6.18 h. Both methods correctly identified 6 patients (in the 100 random sample of charts reviewed) who had been hospitalized for DKA and the dates of their hospital admission. PEPPER not only reduced the time required per chart but also demonstrated substantially lower variability compared with manual review (Fig. 2A). In addition, PEPPER demonstrated distinct groupings of measured times depending on the outcome of each chart review (Fig. 2B). There was a weakly positive correlation between the time taken to review charts manually and the time taken to do the same by PEPPER (r = 0.2517, P < 0.05).

Figure 2.

The time taken per chart and their outcomes. A total of 100 charts were reviewed manually. The same 100 charts were then reviewed by PEPPER. A: Distribution of times taken to review 100 patient charts by PEPPER (blue dots) and manually (orange dots). B: Distribution of times taken to review each chart by PEPPER for criterion I, based on defined characteristics of the charts.

PEPPER identified 110 patients meeting criterion I (Supplementary Fig. 1A). These charts underwent manual review for criteria II–V. In total, 21 patients met full RADIANT criteria for A⁻β⁺ KPD (Supplementary Fig. 1B).

The median age of patients with A⁻β⁺ KPD was 17 years (Q1-Q3 = 15–18). More than half (66.3%) were male, 45.5%, were Hispanic and 40.9% were African American. Non-Hispanic White, Asian, and biracial patients each accounted for 4.5% of the patients with A⁻β⁺ KPD. Patients with A⁻β⁺ KPD had high BMI percentiles (median 98.05; Q1-Q3 = 95.6–99.08). Their median age at diagnosis was 15 years (Q1-Q3 = 12–15). The median HbA_1c at last office visit (median diabetes duration: 15.1 months) was 5.7% (6.5 mmol/L) (Q1-Q3 = 5.2–6% [5.7–7 mmol/L]), and the median C-peptide level (n = 14 of 21 at diagnosis, and n = 7 of 21 at median diabetes duration of 7.2 months) was 1.68 ng/mL (Q1-Q3 = 1.43–2.28) (Table 2).

Table 2.

Characteristics of identified patients meeting full RADIANT A⁻β⁺ KPD criteria

Characteristic	Data
No. of participants with A−β+ KPD	21
Age, median (Q1-Q3), years	17 (15–18)
Sex, n (%)
Female	7 (33.3)
Male	14 (66.6)
Race or ethnicity, n (%)
Non-Hispanic White	1 (4.5)
Hispanic	10 (45.5)
African American/Black	9 (40.9)
Asian	1 (4.5)
Native American/Pacific Islander	0 (0)
Biracial	1 (4.5)
BMI percentile, median (Q1-Q3), kg/m2	98.05 (95.60–99.08)
Positive antibody status, n (%)	0 (0)
Age at diagnosis, median (Q1-Q3), years	15 (12–15)
HbA1c at last office visit, median (Q1-Q3), %	5.7 (5.2–6)
Diabetes duration at last office visit, median (Q1-Q3), months	15.1 (9.3–19.2)
C-peptide, median (Q1-Q3), ng/mL	1.68 (1.43–2.28)
Diabetes duration at C-peptide measurement if not completed at diagnosis, median (Q1-Q3), months*	7.2 (4.2–11.5)

*Of 21 participants, 14 had C-peptide measurement completed at the time of diabetes diagnosis.

Conclusions

We developed and validated PEPPER as a chart review tool capable of automating one specific labor-intensive step of EMR review. In this proof-of-concept application, PEPPER reduced review time by approximately half while maintaining agreement with manual review in the validation subset.

Existing EMR search methods depend on diagnostic codes (6), preexisting structured databases (7–10), manual extraction of data to feed into natural language processing programs (11–13), or existing backend integration into EMRs (14). These limit their usefulness for identifying novel phenotypes when such conditions are not present. PEPPER automates chart access and extraction of prespecified variables directly from the EMR interface without requiring backend data-warehouse integration.

Complex admission histories are less uniformly reported in EMRs. Furthermore, novel or emerging diseases of interest, such as A⁻β⁺ KPD, do not have an established diagnostic code. Our program not only found whether episodes of DKA had occurred in a patient, it also searched the chart by scrolling and clicking to find the earliest incidence of hospitalization for DKA.

PEPPER was used in a retrospective research setting that required the same chart review task to be repeated across large cohorts. In such settings, even partial automation, such as was done here, can meaningfully reduce the overall burden of manual effort and standardize chart navigation across cases.

A strength of this study is that PEPPER was tested in a real-world research use case with direct comparison against manual review. However, PEPPER automated only the first screening step, not full phenotyping. Accuracy was assessed against manual review by a single reviewer, without interrater reliability assessment. Timing comparisons were performed in only the initial 100-chart subset. In addition, PEPPER was configured for one Epic-based workflow and has not yet been validated across other EMR systems, or institutions. Its performance is dependent on the quality of documentation. This version also lacks a user-friendly graphical interface, though one is under development.

In conclusion, PEPPER is a Python-based tool that can automate specific chart-navigation tasks within the EMR and reduce the time required for retrospective chart screening. In this study, it efficiently identified youth who experienced DKA within 6 months of T2D diagnosis, streamlining the first step of identifying A⁻β⁺ KPD candidates. Although broader generalizability remains to be established, PEPPER may be useful for repetitive, discovery-oriented research tasks in which relevant data are not easily retrievable through standard code-based EMR queries or in EMRs lacking natural language processing integration. PEPPER may eventually support clinical workflow by enabling identification of clinically meaningful phenotypes that are not captured by existing ICD codes.

This article contains supplementary material online at https://doi.org/10.2337/figshare.32569341.

Funding Statement

This work was supported by the National Institute of Diabetes and Digestive and Kidney Diseases (grants U54 DK118638, U54 DK118612, and K23 DK129821).