LightGBM estimates treatment-machine leaf errors in separate patient plans
Source check: Source checked. Not independent scientific confirmation.
Finding: A LightGBM model predicted logged multi-leaf collimator position errors in radiotherapy plans from previously unseen patients. Six held-out plans had mean absolute error 0.054 mm; nine additional patient plans supported dose-recalculation agreement. These are machine-log and calculated-dose validation results, with no measured patient benefit or independent physical dose validation.
Technology: LightGBM gradient-boosted decision trees
Publication: Peer-reviewed paper · Test context: retrospective or benchmark. Catalog categories, not an evidence grade.
- AI contribution
- LightGBM learned plan/mechanical/initialization features to estimate delivered leaf-position errors; predicted positions enabled patient-plan dose recalculations compared with log-derived delivery estimates.
- Test and result
- Twenty-nine patient plans were split by patient/plan into 23 development and six held-out test plans. GroupKFold tuning and development-only feature selection preceded final fitting. Against delivered-position machine logs, test MAE was 0.054 mm and RMSE 0.144 mm. Nine further patients supplied independent plans over six fractions; predicted and logged leaf positions were separately inserted into plans for dose recalculation on patient CTs. Agreement concerns reconstructed dose distributions, not in-vivo measured dose.
- Limitation / still unknown
- Single treatment center and accelerator; many leaf/control-point observations derive from few patients. Machine logs are the position reference and both dose distributions use the same dose-calculation system, so no independent physical dosimetry or outcome benefit is established. Dosimetric deviations were small and gamma agreement already high for delivered plans; the study does not establish detection of clinically dangerous failures or safer treatment.
Evidence sources and annotation context
Evidence context (descriptions follow the cited sources; test type is a catalog classification):
Evidence sources: https://pmc.ncbi.nlm.nih.gov/articles/PMC13714354/
Source checked · editorial details
The source page was checked against this entry. Independent confirmation is not assessed by this label.
Source type: paper. Paper: not linked in this catalog; this does not mean no paper exists.
Review date: 21 Nov 2026. Review scope: Read actual primary patient/plan partition, signed-error/log reference, LightGBM tuning/feature selection, separate nine-patient dosimetric Methods, Results and limitations. Original article-meta online October27 is distinguished from November collection date. Supplementary raw logs were not independently inspected. Exact primary identity/date and fresh primary PubMed correction/retraction relationships checked..
No follow-up recorded; this is not a claim that no follow-up exists. Follow-up review scope and date are not recorded. No source-supported relationships recorded.
Editorial assessment: Impact 1/5 · Validation 2/5 · Novelty 2/5. Impact and novelty are editorial judgments; validation follows the annotated test type where one exists. Not an evidence grade. Scoring criteria.