Why calibration matters
A diagnostic classifier can rank cases correctly while still expressing the wrong level of confidence. In health settings, that gap matters: a predicted probability may influence triage, further testing, or clinical attention. Model calibration asks whether predictions labelled, for example, as 70% likely are borne out at roughly that rate.
Research question: how do spline, beta, and Platt calibration scaling compare when used to improve probability estimates in machine-learning disease diagnosis?
The work
The chapter was co-authored by Jeremiah M. Adepoju, Enoch T. Adetunji, and O. Olawale Awe. It evaluates established calibration approaches comparatively, providing evidence about their behaviour in a disease-diagnosis modelling context.
My contribution sits within the collaborative research process. The work compares the methods; it does not claim to have invented spline, beta, or Platt scaling.
Analytical value
- Centres the quality of probability estimates, an important dimension beyond discrimination or headline accuracy.
- Uses comparative evaluation to show that model selection should consider calibration behaviour and intended decision context.
- Supports more responsible interpretation of predictive outputs in a high-consequence domain.
Publication
Predictive Modeling for Disease Diagnosis Using Calibrated Machine Learning: A Comparative Analysis of Spline, Beta, and Platt Calibration Scaling (2025).