KEY POINTS
- Investigators retrieved radiotherapy planning and treatment data collected at the US National Institutes of Health between 2009 and 2024. The export contained 2,482 unique patient identifiers, 4,090 courses, 6,946 plans, and 37,596 structures.
- The original dose-volume histogram database contained approximately 102 million rows and required extensive filtering to distinguish patients, delivered courses, quality-assurance plans, replans, and unused comparison plans. Historical inconsistencies included nonstandard structure names, incomplete clinical annotations, and unreliable plan-intent fields.
- The team developed rt4ml, an open-source Python package that converts dose-volume pairs into patient-level machine-learning datasets. It calculates Dmean, Dmax, Dmin, custom Dx and Vx metrics, biologically effective dose, equivalent dose in 2 Gy fractions, and aggregate dose-volume histograms.
- A pilot analysis included 510 patients with primary intracranial tumors, representing 715 plans across 510 initial treatment courses. Detailed diagnosis and molecular information were available for only 191 patients, illustrating the difficulty of linking historical treatment-planning data with clinical records.
- Machine-learning models using radiotherapy data from all 510 patients achieved accuracies ranging from 0.806 to 0.887 for predicting whether a patient later received reirradiation. However, the highest mean area under the receiver operating characteristic curve was only 0.668, indicating limited separation between patients who did and did not undergo reirradiation.
- Combining clinical and radiotherapy information did not clearly outperform the larger radiotherapy-only dataset. This suggests that increased sample size may have been more useful than adding manually curated variables from a much smaller subgroup.
- Important model features included molecular classification, patient age, brainstem dose, histology, and radiotherapy technique. These associations describe institutional treatment patterns and should not be interpreted as causal criteria for selecting patients for reirradiation.
CLINICAL TAKEAWAY
Historical treatment-planning systems contain large amounts of clinically useful information, but those data are not automatically ready for artificial intelligence or outcomes research. The open-source workflow may help institutions structure dose-volume data at scale, while the modest pilot performance shows that data quantity cannot compensate for inconsistent labels, missing outcomes, and absent external validation.