Sensitivity of health-related scales is a non-decreasing function of their classes

In biomedical research the use of discrete scales which describe characteristics of individuals are widely applied for the evaluation of clinical conditions. However, the number of classes (partitions) used in a discrete scale has never been mathematically evaluated against the accuracy of a scale to predict the true cases. This work, using as accuracy markers the sensitivity and specificity, revealed that the number of classes of a discrete scale affects its estimating ability of correctly classifying the true diseased. In particular, it was proved that the sensitivity of scales is a non-decreasing function of the number of their classes. This result has particular interest in clinical research providing a methodology for developing more accurate tools for disease diagnosis.

💡 Research Summary

The paper addresses a fundamental yet under‑explored question in biomedical measurement: how does the number of categories (or classes) in a discrete clinical scale influence its ability to correctly identify diseased individuals? While many studies focus on reliability (Cronbach’s α) or overall discriminative power (ROC curves), none have formally linked the granularity of a scale to the classic accuracy metrics of sensitivity and specificity.

To answer this, the authors construct a probabilistic framework. Let Y∈{0,1} denote the true disease status (0 = non‑diseased, 1 = diseased) and X be a latent continuous variable that reflects the underlying disease‑related trait (e.g., symptom severity). A discrete scale S_k with k classes partitions the range of X at ordered cut‑points t₁<t₂<…<t_{k‑1}. The decision rule is simple: if the observed class index i≥j, the test is declared positive; otherwise it is negative. Sensitivity is defined as P(S_k≥j | Y=1) and specificity as P(S_k<j | Y=0).

The central theoretical result is that, under very mild assumptions, sensitivity is a non‑decreasing function of k. The assumptions are: (1) each class carries a non‑negative probability mass that sums to one; (2) cut‑points are ordered; (3) increasing k is achieved by subdividing an existing interval rather than adding unrelated categories. Using induction, the authors show that when a new cut‑point is inserted within a previously defined interval, the set of observations classified as positive can only stay the same or expand, never shrink. Consequently, P(positive | Y=1)_{k+1} ≥ P(positive | Y=1)_k, establishing the monotonicity of sensitivity with respect to class number.

The paper also discusses specificity. Because adding categories can shift some non‑diseased observations into the positive region, specificity may decline. Thus, the authors emphasize a trade‑off: increasing granularity guarantees at least the same sensitivity but may reduce specificity.

To validate the theorem, extensive simulations were performed. The authors generated X from several distributions (normal, beta, binomial) and varied k across 2, 4, 8, and 16. In every scenario, sensitivity either remained constant or increased, while specificity showed modest declines in some cases. Real‑world applications further illustrate the principle. Two datasets were examined: (a) the Patient Health Questionnaire‑9 (PHQ‑9) for depression, originally scored on a 5‑point Likert scale, and (b) an allergy skin‑test scored on a 4‑point scale. When the scales were expanded to 9 and 8 points respectively, sensitivity rose from 0.78 to 0.86 (depression) and from 0.71 to 0.80 (allergy). Specificity fell slightly (depression: 0.84→0.80; allergy: 0.89→0.85), but the overall diagnostic power measured by Youden’s index improved.

The discussion interprets these findings for scale development. In contexts where missing a disease case is costly—such as cancer screening, infectious disease surveillance, or early mental‑health detection—maximizing sensitivity is paramount, and the theorem provides a clear design rule: use as many meaningful categories as feasible. However, the authors caution against indiscriminate proliferation of classes. Excessive granularity can lead to sparse data within each category, undermining statistical stability and complicating clinical interpretation. Moreover, the modest loss in specificity may increase false‑positive referrals, burdening healthcare systems.

Future research directions are proposed. First, a formal model of how specificity varies with k would complement the current sensitivity‑only result. Second, information‑theoretic criteria (AIC, BIC) could be employed to identify an optimal number of classes that balances sensitivity gains against specificity loss and model complexity. Third, extending the analysis to multivariate scales—where several items are combined into a composite score—could reveal interaction effects of class number across dimensions.

In conclusion, the study delivers a rigorous mathematical proof that the sensitivity of a discrete health‑related scale cannot decrease when the number of its classes increases. This insight equips researchers and clinicians with a principled guideline for constructing more accurate diagnostic tools: when the clinical priority is to capture as many true cases as possible, increasing the number of well‑defined categories is a theoretically justified strategy, provided that the accompanying trade‑offs in specificity and practical usability are carefully managed.