Overview:
Evaluated machine learning models for predicting stroke risk in adult primary care patients using EMR data. The project aimed to support early intervention and prevention strategies by identifying high-risk individuals automatically within clinical workflows.

​

Tools & Technologies:

Google Colab, python, scikit-learn, pandas, numpy, matplotlib, SHAP, logistic regression, random forest, XGBoost, Neural Network, Naive Bayes 

​​

Background:

Stroke is a leading cause of mortality and long-term disability worldwide, imposing a significant burden on patients, healthcare systems, and economies (Alanazi et al., 2021). In Canada, stroke accounts for an estimated $3.6 billion in yearly costs, nearly 14,000 annual

deaths, and is a major driver of healthcare utilization (Connected by the Numbers, 2023). According to recent research, while stroke deaths have declined, stroke incidence is rising among younger adults, highlighting the need for better prevention strategies (Scott et al., 2022). Early identification of individuals at heightened risk is critical for timely intervention and prevention strategies. Traditional risk assessment tools, such as the Framingham Stroke Risk Profile, provide estimates of stroke probability but often lack the granularity needed for individualized patient management (Chahine et al., 2023). Additionally, traditional methods of risk assessment often require manual calculation by providers, limiting their widespread adoption. Advancements in machine learning (ML) offer an opportunity to enhance stroke risk prediction to develop more precise, personalized and automated risk assessments.

​​

Research Questions:

​

1. For an adult primary care patient in Canada, can we calculate a ‘stroke risk indicator’ (probability of stroke) based on commonly collected primary care data (e.g., age, sex, diagnoses, laboratory values, etc.)?

​

2. Which modifiable health factor(s) offer the biggest opportunity to reduce a patient’s stroke risk?

​​

Rationale:

Predicting stroke risk using primary care data is essential for proactive management. Many risk factors for stroke, including hypertension, diabetes, dyslipidemia, and smoking, are already documented in electronic medical records (EMRs), yet they are often underutilized for

real-time risk stratification. Further, the primary care setting is the ideal care environment for screening and advanced mitigation strategies. An ML-based stroke risk model could allow primary care providers to automatically synthesize the multiplicity of patient-level data in the EMR and systematically identify high-risk patients, enabling targeted interventions. Additionally, understanding which modifiable risk factors have the most significant impact on stroke prevention would enable clinicians to prioritize the most effective treatments.

​

This project aimed to develop a ML-based stroke risk prediction model that could be deployed within a primary care EMR as a clinical decision support system (CDSS). In implementation, if a patient's ‘stroke risk indicator’ surpasses a threshold for stroke likelihood,

an automatic alert would notify the primary care provider, prompting increased vigilance in risk management. This could include intensified blood pressure and lipid control, lifestyle counseling, referral to specialists, and focused patient education on recognizing early stroke symptoms. By integrating such a system into routine care, primary care providers could intervene earlier, reducing stroke incidence and improving long-term patient outcomes.

​

Dataset:

The dataset, obtained from Toronto Metropolitan University in collaboration with the University of Toronto, comprises primary care records collected from 1998 to 2015 across Canada and is considered broadly representative of the national population. Initially, it contained

over 800,000 observations and 175 features (e.g., demographics, clinical measurements, medications, comorbidities), with 58% female patients, 42% male, and ages ranging from 18 to 90.

​

Methods:

After removing approximately 706,000 records that were missing the target variable “stroke,” about 100,000 observations remained. Though there was significant loss of data, the amount of retained observations was considered acceptable for analysis. The target label, defined as a binary variable, had significant class imbalance as expected, with stroke patients representing only 5% (coded as “1”). Features were reduced according to clinical relevance, retaining age, sex, BMI, key lab results, and relevant comorbidities. Comorbidities (0 or 1 binary) were further reduced according to which features for the ‘stroke == 1’ class had a mean of greater than .05 (> 5%), which was selected as an initial cutoff for feature inclusion in predicting the stroke class. Features were also reduced to limit clinical redundancy. The resulting, smaller dataset was saved separately to make further data handling more computationally manageable.

During initial data exploration, four features—LDL, HDL, hemoglobin A1c, and triglycerides—were found to have missing values. Additionally, there was unequal patient representation within the dataset, with some patients having over 100 observations and others only a single observation. The first data cleaning step was to ensure each patient was represented uniformly by handling the multiple observations per patient. A decision was made to use the last observation for a given patient due to the aggregation of common patient records being found to be computationally demanding and the final observation being considered more representative of a patient’s final status compared with the first observation. This step filtered what was initially over 100,000 observations to about 17,000. Even after filtering, the dataset remained highly imbalanced, with 16,626 observations labeled as “no stroke” versus 620 labeled “stroke.” The only categorical variable requiring re-encoding was sex, which was converted to a binary variable (0 = female, 1 = male).

Next, Scatterplots, boxplots, and summary statistics were used to examine continuous features, identify outliers, and remove values deemed clinically implausible or erroneous. Patient ID was removed due to not being predictive. Hemoglobin A1c was removed due to having over 50% missing values and being redundant with diagnosis of ‘Diabetes’. The remaining three features with missingness (LDL, HDL and TG) had a low percentage missing (< 5%) and were retained for imputation. Additionally, a correlation matrix was used to reduce colinear features above a threshold of .9, although no features were found to need removal. This resulted in the final cleaned data for modeling.

A range of models were selected for machine learning analysis to represent a diverse set of learning architectures: Logistic Regression to represent linear models, Random Forest and XGBoost to represent tree-based models, Neural Network to represent deep learning, and Naïve Bayes to represent generative probabilistic classifiers. While all of these models could predict binary outcomes (stroke vs. no stroke), this iterative approach was to determine which model comprised the best balance of performance, explainability, and computational cost. Each model began with a common pipeline step of setting the target label, splitting data into 80% training and 20% testing sets, using MICE imputation for the three features with missing values (to maintain relationships between variables) and scaling continuous features.

To handle class imbalance, each of the models utilized synthetic sampling techniques. SMOTE was used for oversampling the minority class in all models except Logistic Regression, which was found to benefit more from SMOTE-ENN for a combination of oversampling and under sampling. Stratified k-fold cross validation (k=5) was used to preserve class representation in each validation fold and prevent overfitting. Additionally, each model used GridSearchCV for hyperparameter tuning to limit overfitting, for example using regularization in logistic regression and adjusting tree depth in random forest and XGBoost. For the Naïve Bayes model, a hybrid approaches with Gaussian and Bernoulli methods were used to fit a combination of continuous and binary features. For each model run, performance was logged according to metrics of  Accuracy, Precision, Recall, and F-1. Additionally, computational cost was calculated for each model based upon the time required for computation and the size of GPU disk space utilized.

​

Results: