The BAD Paradox: A Critical Assessment of the Belin/Ambr\'osio Deviation Model

TL;DR

This paper critically examines the BAD model for diagnosing keratoconus, revealing systematic biases and inconsistencies that undermine its interpretability and suggesting the need for recalibration to improve clinical reliability.

Contribution

It provides a detailed statistical analysis of the BAD model, identifying sources of bias and proposing improvements for its diagnostic accuracy.

Findings

Systematic bias affects the $D_{final}$ values.

Multicollinearity among predictors complicates interpretation.

Inconsistencies in normative datasets impact model reliability.

Abstract

The Belin/Ambr\'osio Deviation (BAD) model is a widely used diagnostic tool for detecting keratoconus and corneal ectasia. The input to the model is a set of z-score normalized $D$ indices that represent physical characteristics of the cornea. Paradoxically, the output of the model, Total Deviation Value ($D_{\text{final}}$), is reported in standard deviations from the mean, but $D_{\text{final}}$ does not behave like a z-score normalized value. Although thresholds like $D_{\text{final}} \ge 1.6$ for "suspicious" and $D_{\text{final}} \ge 3.0$ for "abnormal" are commonly cited, there is little explanation on how to interpret values outside of those thresholds or to understand how they relate to physical characteristics of the cornea. This study explores the reasons for $D_{\text{final}}$'s apparent inconsistency through a meta-analysis of published data and a more detailed statistical analysis of over 1,600 Pentacam exams. The results reveal that systematic bias in the BAD regression model, multicollinearity among predictors, and inconsistencies in normative datasets contribute to the non-zero mean of $D_{\text{final}}$, complicating its clinical interpretation. These findings highlight critical limitations in the model's design and underscore the need for recalibration to enhance its transparency and diagnostic reliability.