Virtual clinical trial based on outcome modeling with iteratively redistributed extrapolation data.

Area of Science:

• Oncology
• Medical Informatics
• Machine Learning

Background:

• Virtual clinical trials (VCTs) offer computational simulation of clinical studies.
• Limited historical patient data can bias VCT population estimations.
• Accurate prediction of treatment response is crucial for patient outcomes.

Purpose of the Study:

• To develop and evaluate ExMixup, a novel machine learning training technique for VCTs.
• To improve the accuracy of VCTs in predicting cancer recurrence using extrapolated data.
• To assess ExMixup's performance against traditional training methods.

Main Methods:

• Developed ExMixup, a training technique using iteratively redistributed extrapolated data.
• Trained prediction models using original data (baseline), interpolation data (Mixup), and interpolation + extrapolation data (ExMixup).
• Evaluated models using VCTs on prostate and oropharyngeal cancer datasets, predicting radiotherapy recurrence.

Main Results:

• ExMixup models achieved concordance indices of 0.751 (prostate) and 0.752 (oropharyngeal cancer).
• ExMixup significantly outperformed baseline and Mixup models (P < 0.01).
• The VCTs predicted treatment response in patients with distinct characteristics from training data.

Conclusions:

• ExMixup enhances VCTs' ability to predict treatment outcomes, especially for underrepresented patient groups.
• The proposed method addresses data limitations in VCTs by incorporating extrapolated data.
• This approach can improve the reliability of VCTs for diverse patient populations.