Personalized prediction of incident hospitalization for cardiovascular disease in patients with hypertension using machine learning

A retrospective cohort was assembled using linked administrative health databases from Alberta Health with information including demographic and vital statistics, physician billing claims, medication dispensations, hospital separation data, and laboratory services (Fig. 1). These data have been used in previous studies and shown to be high-quality and comprehensive [9, 10].

The study population included all newly diagnosed cases of hypertension aged 18 to 99 years who were residents of Alberta. We identified hypertension cases using a validated case definition of two physician claims within two years or one hospitalization with hypertension related diagnosis codes (ICD-9-CM: 401.x, 402.x, 403.x, 404.x or 405.x; ICD-10-CA: I10.x, I11.x, I12.x, I13.x or I15.x) [11]. The first date of the hypertension diagnosis (index date) was assigned to patients for case definitions with more than one hypertension diagnosis. We included patients who were identified to have hypertension between April 1, 2009, to March 31, 2014 (excluding those who had any code for hypertension between April 1, 2006, to March 31, 2009, thus allowing for a 3-year washout period, and thereby limiting the cohort to only newly diagnosed cases). Patients with prior CVD were also excluded. Cohort assembly is summarized in Fig. 2.

The outcome of interest was incident hospitalization for any CVD, including HF, MI or stroke, identified using validated case definitions: HF (ICD-10-CA: I09.9, I11.0, I13.0, I13.2, I25.5, I42.0, I42.5-I42.9, I43.x, I50.x, P29.0), MI (ICD-10-CA: I21.x, I22.x, I25.2), and stroke (ICD-10-CA: H34.1, I60.x, I61.x, I63.x, I64.x, G45.x) [12]. The time of event was defined as the first occurrence of hospitalization for HF, MI or stroke. We followed patients from initial hypertension diagnosis until the time of the outcome event, death, emigration out of the province, or the end of the study, up to March 31, 2015. If a patient experienced more than one event only the first event was counted as the incident event. Outcome event rates were calculated per 1000 person-years based on a maximum 6-year follow-up period.

Potential predictors were selected a priori based on previous studies and clinical reasoning [13]. Age was categorized into four age groups as the predictor in this study. Patients’ demographic information, such as sex, region of residence was also used as predictors. The number of Charlson comorbidities present in each patient was categorized into “0, 1-2, or ≥ 3.” Fasting blood glucose, estimated glomerular filtration rate (eGFR), cholesterol levels and Glycated haemoglobin (A1c) (HbA1c) was determined between hypertension index date and outcome date [14, 15]. Test results outside the standardized reference intervals were used (Blood glucose > = 7.0 mmol/L, eGFR < 60 mL/min/1.73m², Cholesterol levels > 3.5 mmol/L, HbA1c > =6.5%) [16].

The following categories of antihypertensive medications have been shown to reduce cardiovascular risk and were identified using the anatomical therapeutic chemical (ATC) classification system: beta blockers (ATC codes in category C07, excluding C07AA07, C07AA12 and C07AG02); agents acting on the renin-angiotensin system (ATC codes in category C09); thiazide diuretics (ATC codes in category C03, excluding C03BA08 and C03CA01); calcium channel antagonists (ATC codes in category C08); and miscellaneous antihypertensives (ATC codes in category C02, excluding C02KX01) [17]. Respondents were categorized as using antihypertensive medication if an ATC code corresponded to the above list between the hypertension index date and outcome index date.

The study cohort was randomly divided into training (70% of total: n = 181,911) and validation (30% of total: n = 77,962) sets (Fig. 2). Multicollinearity between predictor variables was assessed using condition indices and variance proportions. Those with significant correlation were removed from the model. The linear multi-task logistic regression (LMTLR) model is an alternative to the Cox’s proportional hazard (CoxPH) model. It can be seen as a series of logistic regression models built on different time intervals to estimate the probability that the event of interest happened within each interval. The constructed LMTLR included 25 features and 50 intervals in this study. The neural multi-task logistic regression (NMTLR) allows the use of Neural Networks within the original multi-task logistic regression (MTLR) design. We used the same 25 features, 100 neurons in the first hidden layer and 100 neurons in the second hidden layer, and one output neuron before input to LMTLR. The random survival forest (RSF) is an extension of the Random Forest model that can take into account censoring individuals. We used 50 trees, the maximum depth of 5 and minimum number of samples required to be at a leaf node at 20 for the model development. The CoxPH is a semi-parametric model that focuses on modeling the hazard function, by assuming that its time component and feature component are proportional over time. The maximum number of iterations in the Newton optimization in this model was 600.

We identified 299,826 newly diagnosed hypertensive patients between April 1, 2009, and March 31, 2015. After applying exclusion criteria, there was a total of 259,873 patients with 899,393 person-years of follow-up and collectively with 11,863 events over a median follow-up time of 3.5 years (inter-quartile range 2.2 to 4.8 years). The incidence rate was 13.4 CVD hospitalizations per 1000 person-years. Among the study population 9182 (3.5%) patients died within the study period. The mortality rate during the follow-up period was 10.0 per 1000 person-years (95% CI: 9.8 to 10.2 per 1000 person-years).

The median age of newly diagnosed hypertension patients was 56.1 years (26.7% were older than 65 years).and 83.6% resided in urban areas. The majority of patients had isolated hypertension without other major comorbidities, but up to one-third had at least one non-cardiovascular Charlson comorbidity, with diabetes being the most common, being present in around 1 in 10 people (9.7%). Nearly two-thirds of patients had at least one laboratory test of interest completed. An elevated LDL-cholesterol (33.7, 95% CI:33.5–33.9), elevated fasting blood glucose (20.5, 95% CI:20.4–20.7), and presence of renal dysfunction (24.5, 95%CI: 24.4–24.7) were the most common laboratory abnormalities. Most patients were dispensed with at least one antihypertensive medication (80.7, 95% CI: 80.6–80.9). Of these, the majority received an angiotensin converting enzyme inhibitor or angiotensin II receptor blocker (67.3, 95% CI: 67.1–67.5), followed by thiazide diuretics (24.7, 95% CI: 24.6–24.9), calcium channel blockers (23.9, 95%CI:23.7–24.1), and beta-blockers (18.1, 95%CI:18.0–18.3) (Table 1, locate at the end of the document text file).

The crude incidence of composite CVD hospitalization was 13.2 (95%CI: 13.0–13.4) per 1000 person-years. Hospitalization for MI was most common (6.1 (95%CI: 6.0–6.3) per 1000 events per person-years), followed by HF (5.6 (95%CI: 5.4–5.7) events per 1000 person-years), and lastly stroke (3.4 (95%CI: 3.3–3.5) events per 1000 person-years) (Table 2, locate at the end of the document text file). The composite CVD hospitalization rate was higher for men, and this was driven by an excess risk of MI. Hospitalizations were most common in patients above the age of 75 years, those residing in rural locations, and individuals with at least two other Charlson comorbidities both for the composite outcome and its individual components.

Figure 3 shows Kaplan-Meier plots of the cumulative probability of being free of hospitalization for any CVD, HF, MI, and stroke as a function of survival time among newly diagnosed hypertension patients. MI had the lowest cumulative probability in the entire survival period when compared with HF and stroke.

Table 3 shows a comparison of the actual and predicted number of hypertension patients experiencing CVD events at each time point for all four models. The RSF model had the lowest RMSE at 33.94 and the lowest MAE at 28.37, indicating the best fit. By plotting the individual’s survival function over time, we compared the survival probability of developing CVD events among individual hypertensive patients. Table 4 shows the results when applying all four models in our hypertension cohort. Overall, all the models had a C-index > 0.70 and Brier score < 0.25 [22], representing a strong predictive ability in the validation set. Adding two layers of neural network before LMTLR, the NMTLR model had the highest C-index (0.82) and lowest Brier score (0.02); the best discrimination and calibration of all the models for event-free survival prediction. This suggests the NMTLR is most accurate and outperformed the other models in predicting CVD outcomes for incident hypertensive patients.

Figure 4 visualizes the LMTLR, NMTLR, RSF and CoxPH models for two representative patients from the validation set. Patient one (red line) had a short event-free survival of 4.0 years from diagnosis of hypertension until being hospitalized for CVD, while patient two (blue line) had a comparatively longer event-free survival of at least 4.9 years before being censored. Patient one developed CVD after 4.0 years diagnosed as hypertension, however, patient two may have been lost to follow-up or did not develop CVD at the end of the study or death until 4.6 years after diagnosed as a hypertensive patient. The median survival time (50% survival probability) as a point estimation for survival time predication was used in the study for personalized survival prediction performance evaluation. If the 50% survival probability is close to the survival time, the model has more accurate prediction performance. All four models performed well in predicting the prognosis for patient one whose 50% survival probability corresponded with the actual observed 4.0 years of event-free survival. However, only the NMTLR model provided accurate prediction of 50% survival probability for patient two who was lost survival information at 4.9 years in this study. For other models, take the LMTLR for example, in fig. 4(a), the survival probability for patient 2 at 4.9 years survival time is near 90%, and this patient’s 50% survival probability is nearby 6.6 years. Although patient 2 passed the 50% survival probability after the 4.9 years in image (a), this patient’s 50% survival probability does not close to the 4.9 survival years, which indicate the model could not well predict this patient’s survival information in this model. The NMTLR was able to handle the presence of censoring better than the other models. Moreover, the individual survival curves for these two patients intercrossed at the beginning of the observation period. This may reflect the real situation that patient two has worse health condition or perhaps the patient is treated and controlled one year after being diagnosed with hypertension. Patient two have a pretty flat curve in the following period, however patient one became worse in the whole follow up period. The CoxPH model was unable to properly handle censoring, as represented by a horizontal survival probability line for patient two.

Figure 5 shows the prediction results for two patients who had CVD outcome at 1.1 years and 2.3 years, respectively. Patient one had a hospitalization for CVD at 2.3 survival years while patient two was had hospitalization for CVD at 1.1 survival years. All of those four models can discriminate the two patients’ survival time versus survival probability. Patient two’s survival curve was always lower than patient one and reached 50% survival probability faster than patient one. For NMTLR, the actual survival time corresponded closely to the estimated survival probability for both patients. For patients two, the survival curve was consistently lower than patient one, and the 50% survival probability occurred earlier for the first patient. Only NMTLR correctly predicted the survival time for patient one at 2.3 years and patient two at 1.1 years, based on the projected 50% survival probability. Moreover, NMTLR also had the smoothest survival curves with distinct shapes predicted for the two patients, while the CoxPH model predicted survival curves with similar shapes because of the proportional hazard assumption. For RSF model, we observed that the survival function was monotonically decreasing and parallel. This is likely due to both patients being in the same tree branch node in the model development process.