The health care industry in the US is one of the largest areas of spending in the country. For 2023, the national health expenditure is expected to top 4.7 trillion dollars, or ~14,000 per person. The US has a complicated payer system in which individuals largely rely on private insurance. Unfortunately, many patients have claims denied which requires extra resources to be assigned to fight these decisions. The state of California permits review of an insurance company’s denied, delayed, or modified service to a patient’s health care plan. This is managed by the California Department of Managed Health Care (DMHC) via an Independent Medical Review (IMR) board. This process occurs when a patients contest a denied, delayed, or modified service by their health care plan. The IMR is carried out by independent physicians with no affiliation to the insurance company. During this process, the patient submits a request for review, and the independent reviewer carefully assesses the medical records, information provided by the patient, their physician, and the insurance company. Subsequently, the reviewer issues a binding decision that is applicable to both parties involved.

How can we develop an accurate and reliable classification model to predict the outcome of the IMR’s decision with the goal of identifying potential biases or disparities in the review process, and ultimately improving the fairness and efficiency of the IMR system for resolving denied, delayed, or modified health care services by insurance companies?

The data was obtained from the California Department of Managed Health Care (DMHC) spanning from 2001 to 2023, which includes over 34,000 IMR decisions. The dataset consisted of both a structured and unstructured portion. The structured data provided information about each claim, such as the report year, diagnosis, treatment details, ruling outcome, gender, and days in the review process. On the other hand, the unstructured data consisted of extensive explanations provided by physicians for each case, averaging around 300 words each, totaling approximately 10 million words across the dataset.