Health claim propensity model using Machine Learning

In health insurance, rising healthcare costs is a concern for both insurer and the insured. In this thesis work we analyze this challenge from the perspective of an insurer. Various factors affect health insurance claims, some of which contribute to formulating insurance policies including specific coverage on critical health ailments. Provisioning of claim payable amount is an important aspect for the insurer. Machine Learning (ML) enables calculating probability of a claim on basis of an insurer will have more accurate forecast visibility on the risk assessment of the insured. To illustrate the performance of each of the ML algorithms that have been leveraged in this thesis work, we analyzed a health survey dataset (BRFSS) available in public domain. Different ML model performances are assessed against metrics such as accuracy, precision, recall and F1-score. The BRFSS dataset is imbalanced in nature. To enable the ML algorithms to generate unbiased output, we leverage balanced approach and adjust the threshold value to generate a balanced trade-off between precision and recall metrics. The final output after balancing the dataset and adjustment of threshold value generated 89% accuracy for LR model, with precision and recall metrics scoring 40% and 39% respectively. The same model when performed on imbalanced dataset, generated similar accuracy, but the precision and recall values were 53% and 13% respectively. Similar results were observed for RF, which although recorded highest precision value (67%) but scored extremely low on recall value (4%). Thus, it is conclusive that implementing balancing techniques on imbalanced datasets and adjustment of threshold value leads to a much more balanced metric analysis and unbiased output.