Source-linked AI summary
A predictive analytics approach for stroke prediction using machine learning and neural networks
Soumyabrata Dev, Hewei Wang, Chidozie Shamrock Nwosu, Nishtha Jain, Bharadwaj Veeravalli, Deepu John
TL;DR
Stroke prediction requires analyzing patients’ electronic health-record attributes and their relative contributions. This paper systematically analyzes those attributes and finds that a perceptron network using four features performs best, with 78% accuracy and a 19% miss rate.
Problem
Existing work had not attempted to analyze the interdependency and relative contribution of risk factors in patients’ health records for stroke prediction.
Method
The paper systematically analyzes electronic health-record attributes using statistical techniques, principal component analysis, and feature-selection methods to identify an optimum feature set.
Results
78% accuracy and 19% miss rate were achieved by a perceptron network using age, heart disease, hypertension, and average glucose level.
Takeaways & Limitations
A perceptron network using the four-feature combination of age, heart disease, hypertension, and average glucose level provides the paper’s best stroke-prediction performance.
Takeaways & Limitations
The perceptron model’s performance could not be improved further because of a lack of additional discriminatory features and additional data.
Abstract
from arXiv · showhide
The negative impact of stroke in society has led to concerted efforts to improve the management and diagnosis of stroke. With an increased synergy between technology and medical diagnosis, caregivers create opportunities for better patient management by systematically mining and archiving the patients' medical records. Therefore, it is vital to study the interdependency of these risk factors in patients' health records and understand their relative contribution to stroke prediction. This paper systematically analyzes the various factors in electronic health records for effective stroke prediction. Using various statistical techniques and principal component analysis, we identify the most important factors for stroke prediction. We conclude that age, heart disease, average glucose level, and hypertension are the most important factors for detecting stroke in patients. Furthermore, a perceptron neural network using these four attributes provides the highest accuracy rate and lowest miss rate compared to using all available input features and other benchmarking algorithms. As the dataset is highly imbalanced concerning the occurrence of stroke, we report our results on a balanced dataset created via sub-sampling techniques.
1. Introduction
The paper motivates stroke prediction from electronic health records by focusing on interdependent risk factors and feature selection. It analyzes patient records, reduces feature dimensionality, and benchmarks machine-learning classifiers to identify useful predictors.
- Research gap: The study addresses the lack of work analyzing all patient-related conditions to identify key factors for stroke prediction.It positions systematic analysis of patient records as a gap in prior stroke-prediction studies.
- Approach: The authors analyze 29072 publicly available patient records containing electronic-health-record attributes for stroke prediction.The dataset includes 11 input attributes and a binary stroke outcome; the patient identifier is excluded from modeling.
- Approach: Principal component analysis transforms the higher-dimensional feature space into a lower-dimensional subspace to assess relative attribute importance.The paper also examines correlations and feature combinations before predictive modeling.
- Evaluation: Several popular machine-learning classification algorithms are benchmarked on the patient-record dataset.The paper compares predictive-modeling approaches after analyzing the available features.
- Motivation: Stroke prediction requires analyzing how patient risk factors relate to one another and affect prediction accuracy.The paper emphasizes interdependency among electronic-health-record factors and their independent effects on predictive performance.
- Dataset: The dataset is highly unbalanced because most records correspond to patients who have not suffered a stroke.This class imbalance is identified as a characteristic of the study data.
2. Analysing Electronic Health Records
The analysis examines feature relationships and predictive importance in electronic health records to identify a compact set of stroke-prediction attributes. Age, heart disease, average glucose level, and hypertension emerge as the most useful features, while CHADS2 scores show increasing stroke occurrence at higher values.
- 2.1. Correlation between features: Correlation analysis finds no highly correlated patient features, suggesting that individual attributes may contribute separately to stroke prediction.Pearson’s correlation coefficient and a color-coded matrix are used to assess relationships among attributes.
- 2.2. Individual Features for stroke prediction: Age has the highest individual importance for predicting stroke, followed by heart disease and average glucose level.Feature importance is measured through the increase in model prediction error caused by each attribute.
- 2.2. Individual Features for stroke prediction: Most EHR observations have low CHADS2 scores, while stroke occurrence increases as the CHADS2 score becomes larger.Most patients with scores of 1 or 2 have a low stroke probability, whereas only a small number with scores above 2 have higher probability.
- 2.3. Selection of Optimum Features for stroke prediction: Adding hypertension to age, heart disease, and average glucose level slightly improves accuracy and slightly decreases the miss rate.Hypertension is identified as the fourth important variable in the feature-importance analysis.
- 2.3. Selection of Optimum Features for stroke prediction: The four features A, HD, AG, and HT are considered optimal because adding other features provides no improvement.Removing age significantly affects accuracy, while removing other features causes no major change.
3. Principal Component Analysis
The paper applies PCA to examine variance, feature relationships, and patient-record representations in stroke-related electronic health records. PCA shows that dimensionality reduction to a few components loses substantial variance, so the analysis retains all components for predictive modelling.
- PCA approach: PCA transforms the 10-dimensional patient-attribute space into linearly uncorrelated principal components that summarize original features.The components are linear combinations of the original feature vectors and lack direct physical interpretation.
- Variance explained: The scree-plot analysis compares variance explained across principal components in the original and balanced datasets.The balanced dataset is created by random sampling to equalize positive and negative stroke samples.
- Feature relationships: In the biplot, age has the largest contribution to the first two components, while average glucose level and heart disease are correlated.Vector length indicates attribute importance, and vector orientation represents relationships in the two-dimensional feature space.
- Variance explained: The first two principal components capture only 31.4% of total input variance, while the first eight explain about 88%.The original dataset requires eight components to explain 88.2% of variance.
- Record representation: The first two components do not adequately represent every health record, indicating that all principal components are needed for good representation.The analysis concludes that feature-space reduction would cause significant loss of variance.
4. Stroke Prediction
The paper benchmarks neural networks, decision trees, random forests, and additional methods for stroke prediction on balanced electronic-health-record data. Neural networks using four selected features achieve strong performance, while principal-component inputs produce comparable accuracy but higher miss rates.
- Benchmark design: The benchmark compares neural network, decision tree, and random forest classifiers using original features, two PCA components, or eight PCA components.Performance is evaluated with precision, recall, F-score, accuracy, miss rate, and fall-out rate over repeated experiments.
- Data preparation: The dataset contains 548 stroke records and 28,524 non-stroke records, creating a highly imbalanced classification problem.The paper addresses this imbalance through random downsampling to 548 minority and 548 majority samples.
- Top-four-feature prediction: 78% accuracy with a 19% miss rate is achieved by a neural network using only age, heart disease, average glucose level, and hypertension.These four features are denoted A, HD, AG, and HT.
- Implications and limitations: The reported four-feature result may support population- or cohort-level resource allocation but is not sufficient to guide individual treatment and prevention.The paper also reports low accuracy variance across repeated downsampling experiments for the benchmarked models.
- PCA-based prediction: Accuracy improves only slightly when all principal components replace the first two, producing similar prediction results with fewer components.The paper links this observation to the importance of variables contributing to the first principal component.
- Prediction results: Neural-network accuracy is comparable between actual features and principal-component inputs, but principal components yield higher miss rates.Because missing stroke cases is important, the analysis favors the four selected features.
5. Conclusion and Future Work
The paper identifies a compact four-feature combination for stroke prediction and reports that a neural network performed best with age, heart disease, hypertension, and average glucose level. It also outlines data and validation limitations and future plans for broader datasets and external validation.
- The analysis systematically examined electronic health-record attributes and their relationships for stroke prediction.
- Age, heart disease, hypertension, and average glucose level formed the selected four-feature combination for stroke prediction.
- Principal component analysis found that nearly all components were needed to explain higher variance, while the first component had the highest variable-loading importance.
- 78% accuracy and 19% miss rate were obtained for the neural network using age, heart disease, hypertension, and average glucose level.
- The perceptron model’s accuracy was constrained by the lack of additional discriminatory features and a larger dataset.
- Future work includes collecting institutional data, integrating background disease and drug knowledge, benchmarking further, and externally validating the proposed method.