Leveraging Machine Learning and Long-Short Term Memory Algorithm for Early Prediction of Diabetes

Abstract

Diabetes, a chronic condition, affects numerous populations. Poor insulin production from the pancreas combined with high blood sugar levels can result in the onset of diabetes. Diabetes can be caused by numerous factors. Observe and prevent these factors to reduce the high prevalence of diabetes. This study concentrates on medical record data for determining diabetes risk factors via statistical correlation analysis. These factors will be utilized as machine learning and LSTM input parameters for diabetes prediction. The factors analyzed include blood glucose levels, HbA1c levels, age, BMI, hypertension, heart disease, smoking habits, and gender. Based on the research results, we found that glucose levels (>137 mg/dL) and HbA1c levels (>6.5%) are the main benchmarks in diagnosing diabetes. It is also supported by the correlation value, which is relatively high (0.42 and 0.40, respectively) compared to other factors. Increasing age and BMI also increase the risk of developing diabetes. Comorbidities (such as hypertension or heart disease) and smoking habits can worsen the condition of people with diabetes. Meanwhile (based on gender), women are more at risk of developing diabetes than men because their body mass index increases during the monthly cycle. Apart from that, there is a tendency for blood sugar levels in women to increase in the last two weeks before menstruation. Based on the prediction results, the highest levels of accuracy, sensitivity, and F1 score were obtained (96.97%, 99.97%, and 98.37%) using the LSTM method. This performance shows that LSTM is relatively good for the diabetes prediction process based on existing factors/parameters.