Reliability of dose volume constraint inference from clinical data.

Area of Science:

• Radiation oncology
• Medical physics
• Biostatistics

Background:

• Dose volume histogram points (DVHPs) are crucial dose constraints in radiotherapy planning.
• Radiation pneumonitis in non-small cell lung cancer (NSCLC) is a relevant clinical context for DVHP studies.
• Logistic regression is a common method for DVHP inference from clinical data.

Purpose of the Study:

• To evaluate the reliability of DVHP inference from clinical radiotherapy data.
• To investigate the impact of cohort size and complication incidence rates on DVHP inference accuracy.
• To compare the performance of Bootstrap and Cohort Replication Monte Carlo (CoRepMC) methods.

Main Methods:

• Generated an 'ideal' cohort from 102 NSCLC dose distributions and a postulated DVHP model.
• Applied Bootstrap and CoRepMC methods to create 1000 equally sized populations for analysis.
• Tested nine scenarios with cohort sizes of 102, 500, and 2000 patients and three DVHP models.
• Analyzed inference frequency distributions and correct inference rates for each scenario.

Main Results:

• Bootstrap method yielded chaotic results for cohort size 1, with improved concentration for sizes 500 and 2000.
• CoRepMC showed increased inference frequency with larger cohort sizes and higher incidence rates.
• Accurate inference rates (>90%) were only achieved with cohorts exceeding 500 patients.
• Both methods indicated high uncertainty in DVHP inference for typical cohort sizes.

Conclusions:

• DVHP inference from clinical data is highly uncertain for standard cohort sizes.
• CoRepMC demonstrated more stable and reliable results compared to Bootstrap.
• Larger patient cohorts (over 500) are necessary for robust DVHP inference in NSCLC radiotherapy.
• Randomness in dose-response significantly influences the statistical analysis of DVHP inference.