Clinical Association Displacement (CAD)

• Clinical Association Displacement (CAD) is a vocabulary-level statistical framework that quantifies shifts in demographic-based word associations in radiology reports.
• It utilizes smoothed log-odds ratios and z-scores to rigorously measure semantic erasure and emergent demographic biases at the term level.
• CAD categorizes shifts into erasure, new bias, bias flip, and preservation, offering actionable insights for clinical fairness audits.

Clinical Association Displacement (CAD) is a vocabulary-level statistical framework designed to quantify shifts in demographic-based word associations in generated radiology reports. CAD detects how model-generated language deviates from reference clinical corpora, focusing on changes in the demographic polarization of clinically meaningful terms. This approach, introduced for the evaluation of vision-LLM (VLM) outputs in radiology, provides a rigorous alternative to surface-level token-overlap metrics by explicitly measuring semantic erasure and emergent demographic biases at the term level (Parikh et al., 2 Mar 2026).

1. Formal Definition and Computation

Let $D\in\{\mathrm{ref},\mathrm{pred}\}$ index reference and predicted corpora of radiology reports, each tagged by a binary demographic variable $g\in\{\mathrm{F},\mathrm{M}\}$. For each term $w$, the group-specific frequency $c_g(w)$ and total count $N_g$ are computed. Dirichlet (add-$\alpha$) smoothing addresses sparse counts:

$p_g(w) = \frac{c_g(w) + \alpha}{N_g + \alpha V},$

where $V$ is the vocabulary size. The log–odds ratio of female to male usage in corpus $D$ is:

$$\hat{s}^{(D)}(w) = \log \left( \frac{p_F(w)}{p_M(w)} \right).$$

Variance is estimated via a multinomial model:

$g\in\{\mathrm{F},\mathrm{M}\}$0

and standardized to a z-score:

$g\in\{\mathrm{F},\mathrm{M}\}$1

Clinical Association Displacement is then defined as the standardized shift between reference and predicted association for each $g\in\{\mathrm{F},\mathrm{M}\}$2:

$g\in\{\mathrm{F},\mathrm{M}\}$3

Statistical significance is determined using the normal approximation ($g\in\{\mathrm{F},\mathrm{M}\}$4 under the null hypothesis), critical threshold $g\in\{\mathrm{F},\mathrm{M}\}$5 (e.g., $g\in\{\mathrm{F},\mathrm{M}\}$6), and Benjamini–Hochberg FDR control over the vocabulary.

2. Statistical and Algorithmic Foundations

CAD is rooted in log-odds ratio association analysis (methodology by Church & Hanks; Monroe et al.), with Dirichlet smoothing to regularize estimates in the presence of rare terms or reports. Frequency uncertainty is accommodated via a delta-method variance approximation under the multinomial, allowing conversion of log-odds to standardized z-scores. Displacement is computed as the difference of independent z-scores, normalized by the root-sum-square of variances, formally yielding a z-test for displacement.

Multiple testing across the vocabulary is accounted for by FDR-corrected $g\in\{\mathrm{F},\mathrm{M}\}$7-values, ensuring robust statistical calibration when flagging significant shifts.

The CAD computation proceeds as follows:

1. Vocabulary is established from all non-stopword tokens in the reference and predicted corpora.
2. For each corpus and demographic group, word counts and smoothed probabilities are computed, leading to group-association z-scores per word.
3. Displacement z-scores are calculated and thresholded for statistical significance.
4. Significant terms are classified as Erasure, New Bias, Bias Flip, or Preservation based on their reference and predicted z-scores.

3. Taxonomy of Detected Shifts

CAD enables the categorization of term-level association changes as follows (with thresholds $g\in\{\mathrm{F},\mathrm{M}\}$8, $g\in\{\mathrm{F},\mathrm{M}\}$9):

Category	Criteria	Interpretation
Erasure	$w$0, $w$1	Previously strong demographic association is eliminated in predictions
New Bias	$w$2, $w$3	Model introduces new demographic association
Bias Flip	$w$4, both $w$5	Polarity of demographic association is reversed
Preservation	Otherwise	Direction and strength of association conserved

This classification reveals clinically salient changes that are not captured by aggregate overlap metrics.

4. Concrete Illustrative Cases

CAD exposes critical pathology in radiology report generation systems that surface metrics (e.g., BLEU) fail to reveal. For example:

• Erasure under Greedy Decoding: The term “mastectomy” shows strong female association in references ($w$6), but is nearly absent in deterministic model outputs ($w$7), resulting in $w$8 and categorization as erasure. Superficial metrics remain high due to template-like predictions, masking this clinical omission.
• Emergent Bias under Stochastic Sampling: A neutral reference association for “pneumothorax” ($w$9) shifts to strong male association in sampled outputs ($c_g(w)$0), $c_g(w)$1—flagged as new bias. Conventional metrics change minutely, failing to detect the emergent demographic correlation (Parikh et al., 2 Mar 2026).

5. Relationship to Weighted Association Erasure (WAE)

While CAD provides term-level specificity, Weighted Association Erasure (WAE) aggregates the $c_g(w)$2 across the vocabulary:

$c_g(w)$3

with the weight $c_g(w)$4 chosen according to analytic emphasis (reference-weighted for importance according to the data distribution, prediction-weighted for model output destabilizations). WAE serves as a global scalar quantifying overall clinical-signal displacement, while CAD pinpoints which terms are specifically suppressed, re-associated, or flipped. The frameworks are complementary and can be deployed in tandem for holistic fairness and fidelity auditing.

6. Practical Implications and Limitations

CAD reveals two modes of model failure in clinical text generation: (1) semantic erasure—where deterministic decoding yields repetitive, generic templates that omit specific clinical terminology—and (2) emergent bias—where stochastic decoding introduces unexpected demographic associations. These are invisible to surface-level metrics, highlighting the inadequacy of current benchmarks for clinical model deployment.

By guiding clinicians and developers towards inference settings that maintain clinical specificity without introducing bias, CAD facilitates more clinically and demographically robust report generation. The framework generalizes to multiple or intersectional demographic axes (e.g., age, ethnicity) if the corpus provides adequate coverage. Limitations include the need for large sample sizes for stable log-odds estimation, sensitivity to pre-processing (stop-word removal), and the requirement for clinical vetting to interpret flagged terms. Post-hoc audit remains necessary to assess the true diagnostic or operational impact of identified shifts (Parikh et al., 2 Mar 2026).

7. Significance for Clinical Model Evaluation

CAD, grounded in smoothed log-odds and rigorous z-testing, addresses a critical blind spot in radiology VLM evaluation by explicitly quantifying what models do not say or say incorrectly in demographic terms. This supports a redefinition of "optimal" reporting beyond token-level similarity, enhancing the detection of semantic omissions and unintended bias. Used alongside WAE, CAD enriches the methodological toolkit for clinical-fairness audits of medical AI systems, providing transparency at both global and granular levels for decision-critical applications (Parikh et al., 2 Mar 2026).