A novel electronic health record-based, machine-learning model to predict severe hypoglycemia leading to hospitalizations in older adults with diabetes: A territory-wide cohort and modeling study

In Hong Kong, the Hospital Authority (HA) established a Big Data Analytics Platform, namely Hospital Authority Data Collaboration Lab (HADCL), to support and facilitate territory-wide secure sharing of EHR-based dataset. HADCL provides anonymized data covering a broad range of patient information collected from all public hospitals and clinics in Hong Kong. The EHR system has provided an integrated, longitudinal, lifelong view of patient’s health status and clinical outcomes, including comprehensive medication and laboratory records, hospitalization, residential area (linked to poverty index), health service utilization, comorbidity, and procedure data [17]. We extracted a retrospective dataset from the HADCL, consisting of patients aged 65 years or above with any in-patient admissions or out-patient attendances from January 1, 2013 to December 31, 2018. All data used was collected for routine patient management with no additional data input required for the modeling [18]. The dataset contains patient demographics, living districts, utilization of health care resources (in-patient admissions, transfer and discharge, out-patient admissions, ED attendance), disease diagnosis based on the International Classification of Diseases Ninth Revision (ICD-9) codes, medication dispensing data, and laboratory tests. Personal information was removed during the analysis procedure. Ethics approval was obtained from the Joint Chinese University of Hong Kong-New Territories East Cluster Clinical Research Ethics Committee (CREC-2021.050).

This study is reported as per transparent reporting of a multivariable prediction model for individual prognosis or diagnosis (TRIPOD) guideline (S1 Checklist).

People with diabetes were defined by those meeting any one of the following criteria [4]: (1) a diagnosis code for diabetes based on ICD-9 code of 250.xx from specialist out-patient clinics (SOPCs) and during hospitalization; (2) diagnosis code for diabetes based on the International Classification of Primary Care, Second Edition (ICPC-2): T89 or T90 at the general out-patient clinics (GOPCs); (3) HbA1c ≥6.5% in any 1 available measurement; (4) fasting plasma glucose (FPG) ≥7.0 mmol/L in any 1 available laboratory measurement; (5) prescription of any glucose-lowering drugs (GLDs); or (6) long-term prescription of insulin for at least 28 consecutive days.

The current study adopted a case-control design. The primary outcome for the cases was hospitalization due to SH, as defined by the principal hospital discharge diagnosis ICD-9 codes (250.80, 250.81, 250.82, 250.83, and 250.30–250.33) [4]. A detailed description of the definition was summarized in S1 Table.

We curated multidimensional variables from the integrated EHR dataset [1,19], which included sociodemographic characteristics, living districts (which are linked to the average income of the residents as an index of social deprivation and poverty), utilization of health care resources (including admissions, clinic visits, consultations by allied healthcare professionals, emergency room visits), disease diagnosis from ICD-9 codes, medication dispensing data, and laboratory data (hematology, renal and liver function tests, glycemic and lipid indexes). These variables were selected based on published literature, prior knowledge, and data availability within EHR. We curated a total of 258 predictors for model derivation. A full list of the predictors and how they were represented in the source data systems and in the prediction models is available in S2 Table.

The prediction horizon (PH) is the time period between model forecasting and the occurrence of a predicted event [20]. In this study, we adopted a PH of 12-month as a balance of SH event rate and clinical utility. To enrich the number of events available for model training, we allowed individuals to have multiple SH events during the whole investigation period.

We defined a 12-month period prior to the date of hospitalization due to SH as the observational period of the cases. For subjects free of SH events, we used the calendar year as the observational periods in controls. We excluded individuals who died in the same year of event onset for cases and during the observational period for controls. We calculated summary statistics for laboratory predictors within the observational periods (mean, median, maximum, and minimum) which are referred to as annual-based values hereinafter. We used 5 consecutive years (2013 to 2017) of data to develop a model that predicted SH events leading to hospitalizations in the subsequent 12 months (2014 to 2018) (Fig 1).

Considering the missingness in EHR data is not at complete random, we additionally defined dummy variables for laboratory predictors as the surrogates of the factors that led to the missingness. We discarded the annual-based values of predictors with missingness >50% and only retained the dummy indicators as the surrogates of these predictors [21]. We imputed the missing values using the cohort median for the remaining features.

We applied 6 supervised ML algorithms for training the SH risk prediction models [22]. These algorithms included generalized linear model (GLM), distributed random forest (DRF), gradient boosting machine (GBM), Rulefit, deep neural network (DNN), and extreme gradient boosting (XGBoost). The whole cohort was randomly split into training (70%), testing (20%), and validation sets (10%) (Fig 1). The models were developed using the training set and optimized via hyper-parameter tuning in the testing set. For benchmarking, we also applied the same 6 ML algorithms to train the models but using only 11 variables that were previously reported to predict hypoglycemia. The best 11-variable model approximates a conventional strategy of SH risk prediction (9–13). These risk factors included age, sex, history of SH, hypertension, blood glucose (HbA1c and FPG), urinary albumin-to-creatinine ratio, estimated-glomerular filtration rate (eGFR) derived from serum creatinine using the Chronic Kidney Disease Epidemiology Collaboration (CKD-EPI) equation, use of metformin, SU, and insulin. We evaluated the models derived from different algorithms and hyper-parameters based on their performance in the validation set. The model development was conducted on the H2O platform (package version: 3.36.1.1) in R environment (www.r-project.org↗) [23].

When training the ML models, we conducted hyper-parameter tuning using either the default setting or a random grid search strategy implemented in the H2O package. In particular, we specified a set of values for the key hyper-parameters that affected learning rate of each ML algorithm (S3 Table). By random grid search, all possible combinations of hyper-parameter values were sampled uniformly from the hyper-parameter space to train the model. We then selected the combination of hyper-parameters that optimized the area under the precision-recall curve (AUPRC) in the testing data for the final models.

We evaluated the model performance using discrimination and calibration metrics. Given the limitations in sole considering AUROC as a discrimination metric, we also considered precision and recall. The former, also known as positive predictive value (PPV), is a measure of the ability of the model to correctly predict a patient as having hypoglycemia, computed by true positive/(true positive + false positive). The latter, known as sensitivity, is a measure of the ability of the model to label as hypoglycemic all of patients who did indeed develop hypoglycemia, signified by the ratio true positives/(true positive + false negatives). The AUPRC was computed at the threshold that yielded the maximum F1 score in the validation set. The F1 score, calculated by the harmonic mean of the precision and recall, measures how well the prediction model can correctly identify all the positive cases and meanwhile avoid making mistakes by marking a negative control as positive. We considered the model with the highest AUPRC as the best. Calibration, the extent to which the predicted risk scores accurately estimate the observed values, was visually assessed by a calibration plot. We compared the observed and predicted risk of SH at 12-month in the validation set by ranking subjects into deciles of predicted risk. In addition, we generated risk probabilities for the outcome event using the best ML model in the training data, and scaled the probabilities to align with a continuous score from 1 to 100 by uniform quantile transformation. We then applied this scaling scheme to the validation set. Score cut-off that enriched 90% of events in the validation data was selected as the threshold for risk stratification.

To assess the performance of the developed model [24], we performed validation in a separate temporal cohort selected from the Hong Kong Diabetes Register (HKDR) [18]. The HKDR is an ongoing prospective register-based cohort of individuals with diabetes since 1995 who have undergone structured diabetes assessments at one of the HA hospitals, Prince of Wales Hospital, Hong Kong SAR. The HKDR cohort was periodically linked to the territory-wide Clinical Management System (CMS) for capturing of laboratory data, treatment, hospitalizations, and death. Characteristics of patients in the HKDR are described elsewhere [18]. This validation cohort was composed of patients with diabetes aged 65 years or above in 2018 and alive by the end of 2019. The same definitions for predictors, outcome, and observational periods were used as in the previous analysis.

We sought to understand how the different variables contributed to the predictions by the XGBoost model (the selected best predictive model). We calculated the variable importance using tree-based algorithms by the H2O platform. The variable importance is computed from the gains of their respective loss functions during tree construction [25]. Additionally, we used Shapley additive explanation (SHAP) value to understand the contribution of each predictor variable in the temporal validation cohort [26].

To interrogate the transferability of our ML model, we additionally performed 2 sensitivity analyses by restricting the predictors to the top 30 variables revealed by variable importance of our XGBoost model. We first re-trained the model using all the 30 predictors using the same training dataset. We then re-trained the second model using 22 predictors that were selected from the top 30 and were considered to be more accessible in routine healthcare and less region-specific. These excluded predictors were mainly outpatient specialty, triage category during ED attendance, ward care type and length of stay during inpatient admission, procedure times, and district of residence. The re-trained models were optimized in the testing cohort and then evaluated in both internal and external temporal validation cohorts.

We presented descriptive statistics as means (standard deviations) or medians (interquartile ranges) to characterize individuals across different years or groups. The ANOVA and χ² tests were employed to compare differences across multiple groups for continuous and categorical variables, respectively.

From January 1, 2013 to December 31, 2017, we identified 1,456,618 patient records of 364,863 individuals with diabetes aged above 65 with valid observational periods (2013 to 2017). The mean age was 74.4 ± 8.0 years and 46.6% of the patients were male. A total of 9,616 unique patients had been hospitalized due to SH during PHs, from which we identified 11,128 outcome events. The prevalence of hospitalization due to SH had declined from 1.3% in 2014 to 0.4% in 2018 (Table 1). Compared with patients without SH hospitalizations, patients who developed SH hospitalizations were older (77.9 ± 7.6 versus 74.4 ± 8.0 years), had more in-patient (3.7 versus 2.1 times/years), out-patient records (18.2 versus 15.1 times/year), and had history of SH (10.0% versus 0.7%). Meanwhile, they were more likely to be taking SU (67.4% versus 43.4%), insulin (48.2% versus 12.3%), and dipeptidyl peptidase-4 inhibitors (DPP-4is) (16.9% versus 7.8%), but were less likely to be taking lipid-regulating drugs (14.9% versus 67.7%) (S4 Table). Distributions of baseline characteristics were similar in training, testing, and validation sets (S4 Table).

All the ML algorithms, including the conventional models using only 11 variables, yielded high AUROC value above 0.8 in training, testing, and validation sets (Tables 2 and S5). Among them, the model based on XGBoost algorithm had the best performance concerning false positives and false negatives in the internal validation set (AUPRC = 0.670; PPV = 0.848). The best ML model based on 11 conventional variables, however, only yielded an AUPRC of 0.280 (XGBoost algorithm; S5 Table). We assessed the model calibration by splitting the validation set into deciles ordered by predicted probability of risk, where the XGBoost-based model demonstrated a good concordance between the observed and predicted events (S1 Fig). We selected a scaled risk probability of 86 as the threshold for risk stratification since approximately 90% of cases were enriched in individuals with scaled scores greater than this cut-off in the validation dataset (S6 Table).

Fig 2A demonstrates the top predictors out of 258 variables and their relative importance in the XGBoost model. The non-use of lipid-regulating drugs, use or historical use of insulin, number of in-patient records, triage category of “urgent” during ED attendance, and use of SU were the top 5 most important variables in the prediction model. Apart from well-established predictors such as medications, age, FPG, and history of SH, the XGBoost model also identified novel variables that could inform the risk of SH including, for example, outpatient appointment specialty, types of wards care, and the district of residence. Sensitivity analysis restricted to the top 30 predictors revealed that the predictive power of our model could be mostly retained by these top predictors (validation AUPRC = 0.632 versus 0.670 for the full-predictor model). Additional sensitivity analysis by further exclusion of region-specific predictors resulted in a 22-variable model with a moderate drop of performance (validation AUPRC = 0.540; S5 Table).

To evaluate the robustness of the XGBoost-based prediction model against the training data collection period (2013 to 2018), we applied the final model to a temporal validation cohort from the HKDR. The HKDR cohort included predictors collected in 2018 and the occurrence of SH hospitalizations in 2019 and included 14,295 valid patients records in 13,917 patients aged 65 or above in 2018, diagnosed with diabetes, and alive by the end of 2019 (Table 1). Using the same outcome definition, we identified 722 SH hospitalizations. The XGBoost-based prediction model yielded an AUROC of 0.856 and an AUPRC of 0.286 in this separate cohort.

Top features revealed by the SHAP value showed high consistency, where non-use of lipid-regulating drugs in the recent 12 months had the largest discriminative power to indicate risk of hospitalizations due to SH. Fewer in-patient records, use of SU and insulin, more out-patient records, more urgent or semi-urgent triage at ED attendance, and lower annual-minimum of FPG were associated with increased risk of SH in subsequent 12 months in this separate validation cohort (Fig 2B).

Data cannot be shared publicly because the raw data is confidential and not available for sharing in accordance with the policy of the Hong Kong Hospital Authority Data Collaboration Lab (HADCL). For inquiries regarding data access, please reach out to HADCL through their website at https://www3.ha.org.hk/data/DCL/↗.