Poster M1 - Efficacy of K-Nearest Neighbors Algorithm for Diabetes Onset Classification

Introduction: Building upon the current understanding of diabetes mellitus classification using machine learning, our research explores the efficacy of the K-Nearest Neighbors (KNN) classification method to enhance prediction accuracy for diabetes onset. Leveraging the studied Pima Indian Diabetes Dataset, which includes diverse physiological metrics from subjects, we aim to refine predictive models by implementing and tuning the KNN algorithm. This approach is motivated by the limitations identified in previous studies where traditional algorithms, while robust, often lack flexibility in handling real-world data variabilities.

Our study particularly addresses the challenge of achieving high classification accuracy amidst the inherent complexities of medical datasets, characterized by imbalanced classes and varying degrees of feature dependencies. The KNN method is chosen for its simplicity and effectiveness in classification tasks where relationships between features significantly influence outcomes. By experimenting with different configurations and feature sets, we optimize our model to capture better the nuanced patterns that signify the progression to diabetes.

Furthermore, our work contributes to the broader field of biomedical engineering by advancing diabetes prediction and enhancing the interpretability of machine learning models in medical diagnostics. In essence, this research aims to refine diabetes prediction models using KNN, offering insights into its applicability and effectiveness in medical diagnostics, and potentially setting a precedent for the adoption of similar techniques in other complex disease prediction scenarios.

Materials and

Methods: The dataset with 768 samples was obtained from the National Institute of Diabetes and Digestive and Kidney Diseases aimed to predict diabetes outcomes, utilizing KNN algorithm, a machine learning method classifying outcomes into binary values. The parameters that were provided in the dataset include Insulin and Glucose levels, Number of Pregnancies, Age, Blood Pressure, BMI, Skin Thickness, Diabetes Pedigree Function, and Diabetes Outcome. Our preliminary analysis showed that Insulin levels, Glucose levels and number of pregnancies were the most viable features for diabetes classification, so we decided to test different combinations of these features to predict diabetes. We tested the KNN classifier using two pairs of features: 1) Insulin levels and number of pregnancies and 2) Insulin levels and Glucose levels. Model training was crucial for the algorithm to discern correlations between variables and diabetes outcomes. Leave-one-out cross-validation method was used to evaluate the trained model's performance. The algorithm's scalar input determined the number of nearest neighbors considered when classifying new data points, assessing the relationship strengths between features like insulin levels and pregnancies.

Cross-validation iteratively tested K values from 1 to100 to optimize model accuracy, sensitivity, and specificity. Contrastingly, other ML techniques, including Naive Bayes, random forest classifiers, and J48 decision trees, are constrained by sample size and assumptions such as axis-parallel decision boundaries, potentially leading to overfitting or underfitting. KNN's adaptability to growing datasets and simplicity post-training sets it apart from these models, as it doesn't require a training phase.

Results, Conclusions, and Discussions: In case 1, the model achieved a classification accuracy of 0.707 at K=13 and specificity of 0.9 (figure 1). Sensitivity was at 0.3. For case 2, classification accuracy was 0.66 at K=59, sensitivity was at 0.95 and specificity was around 0.1 (figure 2). In both cases, accuracy, specificity, or sensitivity plateaued for K>50. The low specificity indicates high false positives showing that the current feature combinations gave us high true positives but less true negatives.

We demonstrated the efficacy of the KNN classification method to predict diabetes. Through our prior analysis of the Pima Indian Diabetes Dataset, we found a significant correlation between physiological metrics and diabetes onset, with insulin levels being the crucial predictor. While KNN offers simplicity and adaptability, its shortcomings is its reliance on the tuning of the function, particularly regarding the number of neighbors. Our analysis shows that the value of K would vary based on the features being used to classify. This is an interesting finding highlighting how different features capture different aspects of the data.

We observed both insulin and glucose levels and insulin levels and number of pregnancies produced similar levels of accuracy. Potential benefits of utilizing this KNN classify method could expand diabetes research, in diagnosing Type 2 diabetes earlier based on factors such as BMI, blood glucose levels, pregnancies, and blood pressure. However, this same method has a level of simplicity that correlates these factors incorrectly when using a lower sensitivity value, like k=2 and misdiagnoses the outcome. Accuracy alone cannot be used to measure the performance of a classifier. From our results both cases showed a high level of sensitivity but a reduced level of specificity. Improvement in these numbers would require bringing in more features that may not necessarily show a correlation with diabetes onset when individually analyzed. For future research, we would like to perform the analysis for more than two features using the KNN classification method and expand this to other populations. Furthermore, this method could also be used to classify other chronic illnesses based on the features included in the data set.

Acknowledgements (Optional):