ML Models Map Lung Cancer Treatment Patterns, But Caution Urged on Clinical Use

Researchers have built machine learning models that can predict what treatment a lung cancer patient received based on their medical history—but they're urging caution before these tools guide actual clinical decisions.

The study analyzed nearly 187,000 patients with non-small cell lung cancer diagnosed between 2010 and 2022, drawing from the Surveillance, Epidemiology, and End Results database, a national cancer registry. The team split the data into two groups: one to develop the models and another, geographically separate, to test whether the predictions held up across different regions of the United States. They trained seven different algorithms to predict whether patients had received surgery, chemotherapy, radiotherapy, or combinations of these treatments.

The models performed impressively on paper. XGBoost, one of the algorithms tested, achieved an internal accuracy score of 0.893 for predicting surgery and 0.853 for chemotherapy—numbers that suggest the system was capturing real patterns in how doctors treat lung cancer. When the researchers tested these same models on the separate geographic validation group, the accuracy remained strong: 0.885 for surgery, 0.835 for chemotherapy. A third algorithm, LightGBM, predicted radiotherapy receipt with a score of 0.766 internally and 0.749 in the validation cohort. Across all models, the strongest predictors of treatment were tumor stage, lymph node involvement, and patient age.

What the models revealed about survival outcomes was more complicated. Patients whose actual treatment matched what the models predicted—those in "concordance" with the algorithm's expectations—showed measurable survival advantages. At the five-year mark, patients who received surgery as predicted lived an average of 15.41 months longer than those who didn't. Chemotherapy concordance was associated with a 9.71-month survival gain. But radiotherapy told a different story. The survival benefit was minimal: just 1.40 months at five years. More troubling, in early-stage disease, patients who received radiotherapy actually showed higher mortality rates—a 56 percent increased risk in stage I and a 32 percent increase in stage II. The survival curves for radiotherapy patients crossed those of non-radiotherapy patients around the 35-month mark, suggesting that any initial benefit reversed over time.

These findings hint at something important: the models are capturing real treatment patterns, but they're not necessarily capturing the reasons behind those patterns. A patient might receive radiotherapy because their cancer is in a location where surgery is impossible, or because they're too frail for chemotherapy. The models can't see those clinical details. They see only the treatment received and the outcome that followed. This is what researchers call "residual confounding"—hidden factors that influence both treatment choice and survival, but aren't visible in the data.

The authors are explicit about the limits of their work. These models should not be used to guide treatment decisions in the clinic, they write, at least not yet. The patterns they've identified are real and reproducible across regions, but they reflect historical practice, not necessarily optimal practice. To move from research tool to clinical tool would require prospective validation—following new patients forward in time, with richer information about their molecular characteristics, the exact timing of treatments, and the reasons clinicians chose one approach over another. Without that, the models risk reinforcing whatever biases or inefficiencies existed in the historical data they learned from. The work is valuable as a demonstration of what machine learning can do in oncology: map complex treatment landscapes with precision. But precision in description is not the same as wisdom in prescription.