Predicting total costs and key drivers in breast cancer surgery patients: ensemble machine learning analyses

Background: As breast cancer continues to present a growing global burden, particularly in China, understanding the factors that drive healthcare costs is crucial for informed policy-making and resource allocation. The primary objective was to identify the key predictors of total hospitalisation costs in breast cancer patients undergoing surgery, using machine learning models. A secondary objective was to explore the influence of different treatment types, patient demographics, and hospital characteristics on total expenses. Methods: We conducted a multicenter, retrospective study utilising an anonymised healthcare dataset collected from 2016 to 2020 across three provinces of Shanxi, Hainan and Liaoning in China. The study included 19,094 breast cancer patients who underwent surgery, identified using the International Classification of Diseases (ICD-10) codes from C50.0 to C50.9 and corresponding mastectomy procedure codes (19301 to 19307). The analysis incorporated a variety of patient characteristics, comorbidities, and hospital attributes. We applied several ensemble machine learning techniques, including gradient boosting algorithms, to assess the contributions of each variable to total costs, both with and without length of stay (LOS). Permutation importance analysis was performed to rank the key cost drivers. A sensitivity analysis using propensity score matching (PSM) adjusted for age, length of stay, insurance type, admission year (2016–2020), week of admission, hospital level (provincial, municipal, district, or other), hospital location, drug fee, and surgery fee was conducted to validate the robustness of the findings, focusing on variables such as drug ratio and tumor surgery admissions. Findings: The average total hospitalisation cost per admission was 2,649.60 USD, with a standard deviation of 2,110.95 USD. LOS was the most significant predictor, with an approximate increase of 150.00 USD per additional hospital day. Other important factors included hospital location, number of beds, and drug ratio. After excluding LOS, the top cost drivers were drug ratio, number of beds, general hospital admissions, tumor surgery admissions, and radiotherapy. Breast cancer patients with longer lengths of stay, admissions to general hospitals in Northern China, a history of radiotherapy, and a lower drug ratio were associated with the highest total costs. The model demonstrated robust performance, with a root mean squared logarithmic error (RMSLE) of 0.474. In the PSM analysis, patients with a drug ratio exceeding 30% had significantly lower average total costs (1,681.65 USD) compared to those with a drug ratio of 30% or lower, who incurred substantially higher costs (2,696.40 USD, P < 0.001). Interpretation: This study underscores the critical role of managing key cost drivers such as LOS and drug ratios in breast cancer surgery. Our results suggest that reducing the duration of hospitalisation and reassessing the allocation of drug costs could lead to lower overall expenses. However, the observed association between higher drug ratios and lower total costs warrants further investigation. Hospital location and capacity were also significant factors, indicating that regional disparities and hospital infrastructure contribute to cost variability. These findings offer valuable insights for healthcare policymakers seeking to enhance cost efficiency and improve patient outcomes in breast cancer care.