Results of an artificial intelligence–based tool to predict financial toxicity early to facilitate early referral to hospital social workers.

Abstract

e13681 Background: Financial toxicity (FT) in cancer care leads to reduced treatment adherence, quality of life, and increased mortality. The Atlanta Cancer Care Foundation (ACCF) provides financial assistance (FA) to all cancer patients through a manual time-consuming application process. All patients referred receive $250, while committee review is required for providing Extra Financial Aid (EFA). AI/ML has not been studied to predict EFA needs in a North American patient population. Methods: Data on 300 patients referred for FA in 2023 was collected, including demographic, zip code, marital status, household size and composition, race, ethnicity, employment, household income, disability status, insurance, goal of requested assistance, diagnosis, requesting agency information, and ACCF adjudication result. A ML/AI model was developed with the above variables using Python, scikit-learn, NumPy, and HistGradientBoostingClassifier. It was trained on 270 patients and tested on 30 separate patients over 20 attempts. The model is open-sourced at Get-Cancer-Aid/ML algorithm at main · Checkmate1721/Get-Cancer-Aid. Chi-square tests were run on the collected variables to identify significant factors predicting EFA. Results: Cancer types requesting FA included breast (35%), GI (18%), heme(13%), lung (10%), gyn-onc, head and neck, GU, other cancer (6% each). Median age was 57 (20-88 years).Median reported income was $21,600 ($0 to $200,000). Race characteristics: black (60%), white (27%), Asian (2%), multiracial (2%), Hispanic (1%), Asian/Pacific Islander (1%), others (7%). Females were 67%, males 33%. Insurance: Private (41%), Medicare (25%), Medicaid (19%), uninsured (10%), VA (1%), no information (4%). All applicants received $250, with 22 patients receiving EFA at committee review. The review process takes an average of 1-2 hours per patient. Chi-square testing showed that Housing Assistance request (P < 0.05) and Private Insurance (P = 0.05) were associated with receiving EFA. The ML model achieved an average prediction accuracy of 90% (range 76%-100% across 20 attempts) in predicting EFA recipients. The model was trained on a diverse set of patients, including minorities, females, rural and urban, and various incomes and payor statuses. Conclusions: AI/ML can effectively predict EFA needs, allowing for prioritized review and quicker assistance. Private insurance and housing assistance requests are associated with receiving EFA. The model's high accuracy suggests potential for streamlining the FA process and improving resource allocation.