Calibration of MRI-based reference intervals to new samples.

Area of Science:

• Neuroimaging
• Biostatistics
• Machine Learning

Background:

• Reference intervals are crucial for interpreting brain magnetic resonance imaging (MRI) findings.
• Existing methods for constructing reference intervals struggle with variations in study design, MRI acquisition, and preprocessing between reference and new datasets.
• Applying MRI reference intervals to new samples presents a significant challenge due to data heterogeneity.

Purpose of the Study:

• To introduce a novel method, Reference interval calibration via conFormal prediction (ReForm), for adjusting MRI reference intervals to new samples.
• To ensure reliable and statistically guaranteed coverage for new observations using conformal prediction principles.
• To address patient privacy concerns by developing a method that does not require sharing of reference data.

Main Methods:

• Developed ReForm, a method building on conformal prediction to calibrate reference intervals for new samples.
• Utilized resampling experiments on Lifespan Brain Chart Consortium cortical thickness data for validation.
• Compared ReForm against refitting intervals, statistical harmonization, and model-based adjustment methods.

Main Results:

• ReForm effectively adjusts reference intervals for new samples, maintaining reliable coverage.
• Empirical results show ReForm controls false positive rates (FPR) comparably or superiorly to methods requiring reference data sharing.
• The method offers a privacy-preserving alternative for calibrating MRI reference intervals.

Conclusions:

• ReForm provides a robust and privacy-conscious solution for applying brain MRI reference intervals to diverse datasets.
• The method demonstrates strong performance in controlling FPR, even without access to original reference data.
• Recommendations for practical implementation and an accompanying R package are provided to facilitate the use of ReForm in clinical neuroimaging research.