Type 1 diabetes (T1D), an immune-mediated chronic disease that affects more than 1 in 300 youth in the United States, is characterized by significant to near-total loss of endogenous insulin production [1,2]. Given insulin’s critical role in maintaining glucose homeostasis, the most immediate and pervasive downstream effect of insulin deficiency is persistent, life-threatening hyperglycemia that must be identified through frequent glucose monitoring and managed with lifelong administration of exogenous insulin [1].

Youth with T1D attend routine (often quarterly) diabetes clinic visits where clinicians use glycated hemoglobin (HbA_1c) testing to assess glycemic status [3,4]. Considered the gold standard for monitoring long-term glycemia in diabetes, HbA_1c testing provides an objective measure of an individual’s mean blood glucose during the previous 2‐3 months [3,5]. To achieve glycemic goals, youth with T1D are increasingly being encouraged to adopt sophisticated diabetes technologies, such as hybrid closed-loop insulin pumps and continuous glucose monitoring (CGM) systems [6,7]. Concurrent with the rising availability of these technologies and a strong research base linking HbA_1c with the development of diabetes complications, the American Diabetes Association has incrementally lowered its recommended HbA_1c goals for youth with diabetes [4,8].

Despite increased adoption of advanced diabetes technologies over time, data from the T1D Exchange indicated that between 2010‐2012 and 2016‐2018, mean HbA_1c in US youth with T1D rose from 7.8% (62 mmol/mol) to 8.4% (68 mmol/mol); and in 2016‐2018, only 16% (686/4346) of youth were meeting the American Diabetes Association’s (then) recommended HbA_1c goal of <7.5% (<58 mmol/mol) [6]. A separate analysis of 2015‐2016 data indicated that fewer than 20% (1817/9685) of US youth with T1D less than the age of 18 years had an HbA_1c<7.5% (58 mmol/mol); and fewer than 10% (690/9685) of youth had an HbA_1c<7% (53 mmol/mol) [9]. Previous research has shown that 1 in 5 youth with T1D experience an increasing HbA_1c trajectory between the ages of 8 and 18 years [10].

Through a phenomenon known as “metabolic memory,” periods of hyperglycemia are known to increase risk for diabetes-related microvascular and macrovascular complications for >10 years following initial exposure [11]. A similar—but beneficial—legacy effect is observed in individuals with T1D who are exposed to near-normal glycemia and later experience more favorable long-term diabetes outcomes, even when glycemic levels later rise [11,12]. These findings point to a critical need to optimize the early identification of youth who are candidates for targeted interventions to improve deteriorating glycemia.

The increasing availability of real-world clinical data housed in electronic health records (EHR) is generating opportunities to investigate population-level health outcomes, develop classification and risk prediction models to augment clinical decision-making, and accelerate diagnostic and therapeutic discovery [13-15]. Machine learning (ML) has been used to meaningfully advance understanding of numerous clinical outcomes in individuals with diabetes [16-18], and EHR-based risk predictions have been leveraged to generate insights across the health-disease spectrum, including T1D [19-21].

Given the multifactorial etiology of rising glycemic levels in youth with T1D, it remains difficult to identify youth who are at the highest risk of experiencing increased HbA_1c between routine diabetes clinic visits. To date, no study has investigated the feasibility of or potential for using EHR data to develop a predictive model to identify youth with T1D who will experience a clinically significant rise in HbA_1c between clinic visits. Such a model could augment clinical decision-making and facilitate initiation of interventions that increase behaviors known to improve glycemia in high-risk youth.

We sought to evaluate the feasibility of using ML to identify youth (aged 9‐18 y) with T1D who were candidates for behavioral and care delivery interventions designed to reduce or prevent a predicted rise in HbA_1c. To do so, we used routinely collected EHR data to develop an interpretable and clinically actionable ML model to forecast unit-change (ie, increase or decrease, in % units) in HbA_1c in 90 days. We then evaluated the ability of our model to augment clinical decision-making by identifying a percent-change cut point that optimized identification of youth who experienced a clinically significant rise in HbA_1c at their subsequent diabetes clinic encounter.

We applied the random forest (RF) regression algorithm to longitudinal EHR data to develop a model to forecast 90-day unit-change in HbA_1c (in % units). We used RF due to its utility for constructing accurate, noise-resilient ML models from high-dimensional data [22,23]. To evaluate our model’s ability to identify youth who, based on predicted rise in HbA_1c, were true candidates for intervention, we evaluated the sensitivity, specificity, positive predictive value (PPV), and negative predictive value (NPV) of predicted versus actual change in HbA_1c at several cut points: ≥0.3%, ≥0.4%, ≥0.5%, and ≥0.6% (approximately 3 mmol/mol, 4 mmol/mol, 5 mmol/mol, and 7 mmol/mol, respectively).

Using data extracted from Oracle Health EHR (formerly Cerner Millenium Electronic Medical Record System; Nashville, Tennessee) [24], we used diagnosis code and laboratory data to identify a cohort of 2757 youth with T1D who received care from a network of pediatric diabetes clinics in the Midwestern United States between January 2012 and August 2017. Criteria used to identify this T1D cohort are provided in Multimedia Appendix 1.

For youth with T1D, we identified health encounters that were associated with HbA_1c measurements (ie, laboratory and point-of-care HbA_1c measurements) and HbA_1c observation windows that met inclusion criteria. Each HbA_1c observation window comprised 2 documented HbA_1c measurements (from a single individual) separated by a time interval of 70‐110 days. The 70‐ to 110-day time interval was selected to approximate the 3-month (ie, 90-day) time interval between regularly scheduled diabetes clinic visits.

Certain encounters with HbA_1c data were excluded from consideration and therefore not included in any HbA_1c observation windows. HbA_1c values documented at or shortly after T1D diagnosis tend to be more extreme than those documented at subsequent time points (ie, after an individual with T1D begins receiving regular insulin injections) [25,26]. As such, each youth’s first-documented encounter with an HbA_1c value was excluded under the assumption that a youth’s first HbA_1c measurement may have been obtained at the time of T1D diagnosis. We also excluded data from encounters where youth were <9 years old, as the incidence of clinically significant rise in HbA_1c is less common in this age group [6,27].

We excluded observation windows associated with HbA_1c measurements that were separated by <70 days or >110 days, as well as those where the first encounter for a given HbA_1c observation window (ie, the index encounter) was associated with an HbA_1c of >12% (>108 mmol/mol). The latter exclusion criterion was used because individuals with an HbA_1c of >12% (>108 mmol/mol) were already considered ideal candidates for intervention. Encounter-level data from all HbA_1c observation windows that met inclusion criteria were included in our final dataset, which could include data from multiple HbA_1c observation windows per individual.

The forecasted outcome was unit-change in HbA_1c (in % units) at the end of 90 days. After predicting each youth’s percent change in HbA_1c in 90 days (ie, at the time of the follow-up encounter), we used various thresholds to determine an HbA_1c percent rise cut point that optimized identification of individuals who were true candidates for intervention at the time of their index encounter: ≥0.3%, ≥0.4%, ≥0.5%, and ≥0.6% (approximately 3 mmol/mol, 4 mmol/mol, 5 mmol/mol, and 7 mmol/mol, respectively). We considered these cut points to be clinically relevant and actionable, given that a long-term decrease of ≥0.3% (3 mmol/mol) in HbA_1c is associated with reduced risk of long-term diabetes complications [28].

We used SQL queries to comprehensively extract up to approximately 5 years (January 2012-August 2017) of structured and unstructured EHR data for each youth with index and follow-up encounter data for at least 1 qualifying HbA_1c observation window. These data included demographics, laboratory results, vital signs, anthropometric measures, encounter locations, medications, diagnosis codes, procedure codes, structured clinical vocabulary codes, and free-text data from diabetes- and non–diabetes-related clinical notes, messages, and reports.

Demographic data included sex (female, male), age, ethnicity (non-Hispanic, Hispanic), race (White, Black or African American, Asian, American Indian or Alaska Native, Native Hawaiian or Pacific Islander, and other), primary language (eg, English or Spanish), health plan type; and postal code (3- and 4-digit postal code prefixes). Additional extracted data included up to approximately 5 years of all available laboratory test results, clinical event and observation data, vital signs (heart rate, respiratory rate, oxygen saturation, and blood pressure), anthropometric measures (weight, height, and BMI), and medications (mapped to standard generic drug names [29]). We also extracted diagnosis codes (ie, ICD-9 [International Classification of Diseases, Ninth Revision], ICD-10 [International Statistical Classification of Diseases, Tenth Revision], and Systematized Nomenclature of Medicine Clinical Terms [SNOMED CT] codes), procedure codes (ie, Current Procedural Terminology [CPT] codes); and other structured clinical vocabulary codes (ie, SNOMED CT).

We chose not to include data generated by diabetes devices (eg, automated insulin delivery and CGM systems). Early on, we observed that HbA_1c was easiest to predict in youth who used diabetes devices that generate diabetes data (eg, glucose levels) in real time. However, since most diabetes centers do not have broad or ready access to device data in near-real time, we sought to evaluate the potential of using only EHR data to predict HbA_1c.

We engineered features using data documented during all available historical encounters, as well as during HbA_1c observation window index and follow-up encounters. Processes used to transform variables into features for ML varied by data type. In all, our feature engineering processes generated 17,466 input features for model fitting.

For numeric variables (eg, laboratory results, weight, and vital signs), we created features by calculating summary metrics (ie, mean, slope, and SD). In general, we created 2 sets of features for each numeric variable, based on proximity of the measurements to the HbA_1c observation window’s index encounter. One set of features was created using data documented during the 12 months preceding (and at) the index encounter. A second set was created using all available EHR data documented before (and at) the index encounter. For example, we created 2 features for mean HbA_1c: one calculated using the previous 12 months of HbA_1c data (up to and including the index encounter), and the other calculated using all available HbA_1c data (up to and including the index encounter). Given the intrinsic insensitivity of RF to numerical outliers, we did not alter or drop outliers from the data. Once all numerical features were created, missing numerical values were imputed using the population median.

Each youth’s diagnostic (ie, first) HbA_1c result was included as a separate feature, as was the HbA_1c result documented at the observation window’s index encounter. Because research suggests that youth with T1D can be grouped into one of several HbA_1c trajectory clusters [10], we created an HbA_1c trajectory feature by using k-means clustering [30] to assign youth to 1 of 4 clusters based on their quarterly HbA_1c measurements.

We used data documented at the observation window’s index encounter to create features from demographic data (eg, age, race, ethnicity, primary language, health plan type, and postal code). For each categorical demographic variable, we used the StringIndexer feature transformer to convert the categories associated with each variable into numeric indices, thus creating a single feature for each of these variables [31].

We used Clinical Classification Software Revised (CCSR), developed by the Agency for Healthcare Research and Quality, to group ICD-10 codes into meaningful categories [32]. Thereafter, each CCSR category and each ICD-9, ICD-10, SNOMED CT, and CPT code was treated as a separate variable. We created 2 sets of features for each of these separate variables, based on how many times each had been assigned to the individual relative to the observation window’s index encounter. One set of features was created by calculating the frequency that each had been assigned to the individual during the 12 months preceding (and at) the index encounter. The second set was created using all available EHR data documented before (and at) the index encounter. Absence of diagnosis, procedure, or structured clinical vocabulary codes was presumed to reflect true absence, rather than missingness, of these data variables.

Medication variables were similarly transformed into 2 sets of features based on how often each medication had been prescribed relative to the index encounter. One set of features was created by calculating the frequency that each medication had been prescribed to the individual during the 12 months preceding (and at) the index encounter. The second set was created using all available medication data documented before (and at) the index encounter. Encounter frequencies were similarly calculated and included as separate features. Absence of medication and encounter data was presumed to reflect true absence of these data.

We used term frequency–inverse document frequency (TF-IDF) vectorization, a natural language processing technique, to process free-text data from clinical notes, messages, and reports. In TF-IDF vectorization, words (ie, tokens) are first converted into a matrix of token counts [33]. The matrix is then transformed into a normalized TF-IDF representation that most heavily weights tokens that occur infrequently across the entire corpus of available text [33]. As such, TF-IDF is used to assign the highest weight to words that have the most discriminating power. After ranking by weight, we constrained the total number of features generated via TF-IDF vectorization to 250 single-word terms and 250 two-word terms, each of which had to be present in at least 5 documents.

RF uses bootstrap aggregation and random feature sampling to independently train a series of uncorrelated decision tree regressors, known as “weak learners” [22,23,34-36]. Predictions from this ensemble of weak learners are averaged to produce a single “strong learner” with improved prediction accuracy [23]. Relative to many other ML methods, the RF algorithm presents several key advantages, including decreased risk of overfitting, straightforward calculation of the degree to which individual input features contribute to model predictions, and robustness to missing data [22,34].

After randomly splitting the entire dataset into 3 nonoverlapping data subsets, we used 3-fold cross-validation to recursively fit RF regressors to 2 of the 3 subsets and then evaluate model performance on the third, held-out subset. We used 3-fold (rather than 5- or 10-fold) cross-validation due to the large number of HbA_1c observation windows included in our analysis, as well as our desire to reduce variance in the estimated performance of our model. Hyperparameters used for model fitting are presented in Table 1. Model performance was evaluated by averaging the mean absolute error (MAE) and the root-mean-square error (RMSE)—the SD of the residuals [37]—across all 3 cross-validation models.

Decision tree regressors are grown by recursively splitting on features to maximize impurity reduction [39]. Feature splits that reduce impurity by maximally reducing variance are considered important; thus, the features that are split to maximize reduction in variance are also deemed important [39,40]. We evaluated feature importance by calculating and ranking the mean reduction in variance associated with only those features that were used by all 3 of our cross-validation models to forecast HbA_1c.

We used Python (version 3) and Scala (version 2; Programming Methods Laboratory at École Polytechnique Fédérale de Lausanne) to clean and transform the data. ML analyses were conducted using the Apache Spark MLlib (version 2) ML library [41].

Pearson r correlations were used to assess the strength and direction of the relationship between actual and predicted HbA_1c values. We also used sensitivity, specificity, PPV, and NPV as clinical performance metrics to aid in identifying a predicted HbA_1c percent rise threshold that would facilitate optimal capture of youth who would experience a clinically significant rise in HbA_1c in 90 days.

Summary statistics, correlations, RMSE, MAE, and sensitivity, specificity, PPV, and NPV metrics were assessed using Stata/SE (Stata standard edition) software (version 18.5; StataCorp) [42].

Clinical and model output data were collected and coded in an institutional review board–approved research data repository at Children’s Mercy Kansas City (Kansas City, Missouri; IRB #11120355) that met the requirements for a waiver of written informed consent as outlined in 45 CFR 46.116.

Out of 2757 youth with T1D, 1743 youth (63.2%) had one or more HbA_1c observation windows (n=9643) that met inclusion criteria (Figure 1).

Characteristics of the entire cohort that met inclusion criteria are summarized in Table 2. The observed frequencies that youth experienced a rise in HbA_1c that exceeded each percent change cut points (≥0.3%, ≥0.4%, ≥0.5%, and ≥0.6% [approximately 3 mmol/mol, 4 mmol/mol, 5 mmol/mol, 7 mmol/mol]) were 40.7%, 35.6%, 30.8%, and 26.5%, respectively. Characteristics of observations included in each nonoverlapping K-fold are summarized in Table 3.

Across all 3 folds of our cross-validation model, average RMSE was 0.88 (Figure 2). Thus, in 68% (6557/9643) of cases (representing one SD), our predictions were within ±0.88% (95% CI 0.85‐0.90) of the true percent change in HbA_1c. The average MAE across all 3 folds was 0.64 (95% CI 0.63‐0.65). Predicted HbA_1c(%) strongly correlated with true HbA_1c(%; r=0.79; 95% CI 0.78‐0.80).

Across all 3 folds of our cross-validation model, the top 10 features identified as having the greatest impact on model predictions included postal code, various metrics related to HbA_1c, and the number of times that the individual had been assigned a diagnosis code indicating difficulty with treatment engagement (Figure 3). The top 30 most important features used to predict percent change in HbA_1c are in Multimedia Appendix 2.

Our cross-validation model’s ability to accurately predict change in HbA_1c at various percent change cut points is illustrated in Table 4. At each percent change cut point (≥0.3%, ≥0.4%, ≥0.5%, and ≥0.6% [approximately 3 mmol/mol, 4 mmol/mol, 5 mmol/mol, 7 mmol/mol]), PPV was 60.3%, 56.4%, 52.7%, and 53.1%, respectively, indicating an approximately 1.5- to 2-fold enrichment (relative to the observed frequency of each outcome [Table 1]) for identifying youth who would experience a clinically significant rise in HbA_1c. Sensitivity and PPV improved when predictions involved smaller changes in HbA_1c, whereas specificity and NPV improved when predictions involved larger changes in HbA_1c. Sensitivity, specificity, PPV, and NPV metrics for each K-fold are in Multimedia Appendix 3.

We used routinely collected EHR data, including both structured and unstructured data, to establish the feasibility of constructing an interpretable ML model for predicting unit-change in HbA_1c (in % units) between quarterly diabetes clinic visits among youth (aged 9‐18 y) with T1D. For those predicted to experience a ≥0.3% (approximately 3 mmol/mol) rise in HbA_1c during the following 3 months, PPV was 60.3%, indicating a 1.5-fold enrichment (relative to the observed frequency [40.7%] of this outcome) for identifying youth who would experience a clinically significant rise in HbA_1c. This finding, which suggests that EHR data may be useful for identifying youth who will experience rising glycemic levels, is clinically relevant given that a long-term increase of ≥0.3% (3 mmol/mol) in HbA_1c is associated with increased risk for long-term complications of diabetes [28].

Another key finding was that our model’s sensitivity and PPV were higher when the predicted percent rise threshold was lower (eg, ≥0.3% vs ≥0.4%), whereas specificity and NPV were increased at higher predicted percent rise thresholds (eg, ≥0.4% vs ≥0.3%). We hypothesized that using a higher percent rise threshold would decrease the likelihood of false positives (ie, identifying a youth as someone who would experience a corresponding rise in HbA_1c when they did not), and the data supported this conclusion. On the other hand, using a lower percent rise threshold reduced the likelihood of missing those who would experience a clinically significant rise in HbA_1c. If confirmed in future studies, these findings suggest that using the lowest clinically significant threshold may be useful for guiding clinical decision-making and subsequent initiation of interventions designed to mitigate rising glycemic levels.

We also evaluated our model’s ability to augment clinical decision-making by using PPV and NPV to identify a percent-change cut point that optimized identification of youth who experienced a clinically significant rise in HbA_1c at their subsequent diabetes clinic encounter. Although PPV and NPV are considered the metrics of choice for clinical decision-making at the level of an individual person, the selection of desirable PPV and NPV values in a particular use case depends on numerous factors. These factors include considerations about short- and long-term burdens and costs related to over- or undertreatment, associated psychological impacts on individuals receiving care, and short- and long-term costs imposed on the health care system (eg, for increased staffing resources) [43]. Therefore, before implementing this model clinically, it would be important to allow clinicians to provide feedback about the most appropriate thresholds for defining clinically significant rise in HbA_1c, along with associated PPV and NPV values. For this work, we propose using the ≥0.3% cut-point to maximize capture of high-risk youth who are candidates for behavioral and care delivery interventions designed to reduce or prevent predicted rise in HbA_1c.

The top features impacting our model’s predictions (ie, postal code, numerous metrics pertaining to HbA_1c, and history of low treatment engagement) have been shown in previous studies to be associated with elevated glycemic levels. Ample evidence suggests associations between geographic location and geographically linked measures of socioeconomic status (eg, area deprivation, social deprivation, and child opportunity indices) and T1D outcomes, including glycemic levels and diabetic ketoacidosis [44-47]. Previous HbA_1c measurements have also been shown to significantly impact ML-based predictions of future HbA_1c, but previous investigations have only examined this in adults with type 2 diabetes (T2D) [48]. Finally, lower treatment engagement has been shown to have a substantial impact on HbA_1c in youth with T1D [49,50]. This evidence collectively underscores the critical need for members of the diabetes care team to partner with affected youth and families to identify resources and tailored strategies for optimizing diabetes self-management behaviors.

Given the widespread use of EHRs in clinical care, as well as the growing volume and availability of these data, there exists tremendous potential for using EHR data to identify and personalize care pathways for improving health outcomes in T1D. Previous work has applied ML to EHR data, for example, to predict the onset of T1D in youth [51], as well as diabetic ketoacidosis in both youth and adults with T1D [19,20,52]. Recent research has focused on applying numerous ML classifiers to medical encounter data to predict HbA_1c in individuals with T2D [48]. The area under the receiver operating curve for each of the top 5 best-performing classifiers in the aforementioned study was extremely high (>0.95). Of note, however, these model predictions were binary (ie, HbA_1c <7% [<53 mmol/mol] vs ≥7% [≥53 mmol/mol]) rather than continuous and were evaluated in a primarily adult Chinese cohort diagnosed with T2D, limiting generalizability to other populations. Our approach is designed to predict unit change in HbA_1c and to give clinicians a simple output (ie, HbA_1c will or will not increase by ≥0.3%) for interpretation. This study is the first to use EHR data to predict a clinically significant rise in HbA_1c in youth with T1D.

Recent efforts have also explored the use of ML classifiers that use 2 weeks of CGM data to forecast 90-day HbA_1c in youth with T1D [53]. The first of these studies used a nested, ensemble learning approach to iteratively predict HbA_1c in stages: (1) HbA_1c ≤7.5% (58 mmol/mol) or >7.5% (stage 1), (2) HbA_1c ≤9% (75 mmol/mol) or >9% (stage 2, after stage 1 was complete), and (3) HbA_1c ≤12.5% (113 mmol/mol) or >12.5% (stage 3, after stage 2 was complete) [54]. A subsequent study used few-shot learning followed by K-nearest neighbors to classify transformed images of CGM time series data into multiclass HbA_1c intervals [55]. Generalizability of these HbA_1c prediction efforts is limited, however, by these methods’ dependence on CGM data and by racial disparities in the relationship between CGM metrics and HbA_1c [56].

Currently, CGM systems are neither accessible to nor used by all individuals with T1D. Recent data from the T1D Exchange Quality Improvement Collaborative suggest that only 40%‐50% of US youth with T1D currently use CGM systems [57,58]. Reasons for this are multifactorial and can include reluctance to use CGM technologies, financial constraints, lack of insurance coverage, device-related skin complications, CGM alarm fatigue, and sociodemographic and racial or ethnic disparities in access that adversely impact use of diabetes technologies [58-62]. At this time, CGM data also remain notably absent from most EHRs, are distributed across multiple proprietary commercial software, and are difficult for health systems to access. Although efforts to integrate CGM data into the EHR remain ongoing [63,64], large-scale implementation of these efforts will hinge on the development of CGM-related data standards and a data architecture that supports this integration [65].

In contrast, EHR data are routinely collected on every person receiving care from a given health care institution. These data thus provide a rich, longitudinal source of individual- and population-level health data that can be leveraged in near real-time for ML-driven clinical decision support [13,66]. Even so, the potential for integrating EHR-based ML-driven analytics in health care remains largely unrealized. A 2020 systematic review evaluating the number of clinical prediction models that have been embedded into EHRs noted that fewer than 45 such examples have been published [67]. Of note, only 36% (16/45) of model implementations occurred in outpatient settings, and none of the embedded models were specific to individuals affected by diabetes [67]. These findings highlight a critical gap, as well as opportunity, for leveraging real-world EHR data to facilitate real-time risk prediction and improve diabetes-related health outcomes.

A strength of this study is its use of longitudinal EHR data to predict 90-day unit-change in HbA_1c in a large cohort of youth with T1D. The scale and granularity of these data facilitated the creation of thousands of data features that we simultaneously analyzed as potential predictors for suboptimal glycemic outcomes. Additional strengths of this study include its use of explainable ML methods for evaluating model predictions and our use of a clinician-led, postmodeling decision analysis to enhance clinicians’ understanding and uptake of model predictions. The relevance of our model is underscored by its ability to forecast 90-day change in HbA_1c for all youth receiving care through our regional clinic network, and not only for those using CGM systems.

Several limitations also warrant consideration. The data used in this study originated from a regional network of diabetes clinics in the Midwest United States and may not generalize to other geographic locations or health care settings, to future cohorts using rapidly evolving diabetes treatment technologies, or to more racially or socioeconomically diverse cohorts. External validation of the geographic and demographic “transportability” of this and future iterations of our model will hinge on ensuring that data from different clinical settings are collected in similar ways and standardized according to a common data model. Examples of such data standards include the Observational Medical Outcomes Partnership Common Data Model [68] and the T1D Exchange Quality Improvement Collaborative data specification [69]. As well, EHR data are subject to data entry errors and missing data that inadvertently occur as a part of routine clinical care. EHRs are also characterized by data fragmentation and reflect biases in clinical data collection, documentation, and decision-making [13]. Therefore, results from this and all models constructed using EHR data must be interpreted carefully, given both known and unknown biases that impact model predictions.

Model generalizability could be enhanced by using standardized geographic-based features (eg, an area deprivation index or the Child Opportunity Index [45,47]) rather than zip code, as well as by creating a final prediction model that includes only a limited number of the “top-N features identified via cross-validation. Using additional data preprocessing methods (eg, one-hot encoding) when transforming categorical demographic features (eg, race and ethnicity) for ML would facilitate interpretability of model results pertaining to those features. Model performance may improve with additional hyperparameter tuning. This model’s predictive utility could also be compared with that of models constructed using other ML methods, including other explainable AI methods and deep learning models. Finally, for youth who adopt diabetes technologies, such as CGM and automated insulin delivery systems, the inclusion of diabetes device data would likely significantly augment our model’s predictions.

We acknowledge that translation of this work into clinical practice will be accompanied by various logistical and practical challenges. This study was designed as an “initial step” to evaluate the feasibility of using EHR data to predict change in HbA_1c. As previously described, additional research is needed to address issues related to model refinement, validation (using data from external organizations, as well as future EHR data collected from our network of diabetes centers), and deployment in clinical settings. Future work can, for example, evaluate whether a limited set of standardized features may be useful for developing a more parsimonious model that can be readily disseminated to other institutions. Once deployed, ongoing monitoring of model performance will also be needed.

Furthermore, we acknowledge that refining and successfully incorporating this approach into clinical and decision workflows will hinge on the collection of additional evidence from future studies with even larger and more diverse patient cohorts, as well as buy-in and trust from both clinicians and patients. Although, in this iteration, our modeling approach yielded a nontrivial number of false positives, we note as well that our model’s performance represents a substantial improvement over existing capabilities. Compared, for example, with initiating interventions randomly or initiating interventions at every diabetes clinic visit (to address youths’ rising glycemic levels, which occurred 40.7% of the time in our cohort), our modeling efforts facilitated pre-emptive identification of rising glucose levels three-fifths of the time. The 1.5-fold risk enrichment demonstrated in this work represents a meaningfully improved opportunity for more targeted initiation and delivery of interventions designed to lower youths’ glucose levels.

Using EHR data to develop an ML-based prediction model to identify youth who will experience a clinically significant rise in HbA_1c between diabetes clinic visits is both timely and feasible. Future research should aim to further optimize model performance, as well as evaluate model performance in racially or ethnically, socioeconomically, and geographically diverse cohorts. Future work is also needed to evaluate whether model results vary by duration of diabetes, use of technology (eg, CGM system users vs nonusers), and insulin delivery modality. Findings from this study may help to inform risk stratification and resource allocation efforts and serve as a catalyst for future quality improvement efforts focused on developing and evaluating personalized strategies and supports for optimizing diabetes self-management behaviors.