Medical Similarity Index Overview

• Medical Similarity Index is a domain-calibrated scoring function that quantifies the similarity between various medical objects through tailored representations and task-specific thresholds.
• It encompasses methods ranging from semantic evaluation in clinical text to morphological analysis in imaging and signal processing, each employing distinct mathematical formulations.
• Its practical applications impact diagnostic accuracy, retrieval in systematic reviews, and patient similarity assessments, emphasizing the need for domain-specific calibration.

Taken as an umbrella term, a Medical Similarity Index denotes a quantitative score that measures how alike two medically relevant objects are. In the cited literature, those objects include medical questions, clinical sentences, abstracts, patient trajectories, drugs, MedDRA terms, beam profiles, segmentation contours, diagnostic-test distributions, and synthetic-image datasets. The term itself is not used uniformly across these works, but the underlying pattern is stable: each method maps a medically structured pair—or, in some cases, two medically structured distributions—to a scalar whose meaning is task-specific but clinically or semantically interpretable (McCreery et al., 2019, Wang et al., 2018, Lei et al., 2018, Fazekas et al., 13 Aug 2025).

1. Scope and recurrent mathematical forms

A Medical Similarity Index is not a single canonical formula. The literature instead instantiates it through several recurring forms. In binary semantic matching, the score is often a posterior probability such as

$\mathrm{MSI}(q_1,q_2)=p(y=1\mid q_1,q_2),$

where $y=1$ denotes “same medical question” (McCreery et al., 2019). In clinical semantic textual similarity, the index may be an ordinal expert score on a $0$–$5$ scale, where $0$ is completely dissimilar and $5$ is completely equivalent (Wang et al., 2018). In diagnostic accuracy, the index can be the Hellinger affinity

$\kappa=\int \sqrt{f_D(y)}\sqrt{f_{\bar D}(y)}\,dy,$

which measures similarity between diseased and non-diseased test-result distributions (Carvalho et al., 2017). In segmentation, it can be the average of point-wise scores derived from bidirectional local distances, with separate penalties for inside and outside contour deviations (Fazekas et al., 13 Aug 2025).

Comparison target	Typical score form	Representative source
Medical questions or texts	Probability, ordinal similarity, cosine-based score	(McCreery et al., 2019, Wang et al., 2018, Picha et al., 2024)
Patients or documents	Distance, bilinear similarity, ranking score	(Zhu et al., 2019, Goyal et al., 2018, Lee et al., 2021)
Signals, images, contours	Correlation-style coefficient, overlap-style score, weighted boundary average	(Bermudez et al., 28 Mar 2025, Eskandari et al., 2023, Fazekas et al., 13 Aug 2025)
Ontology or vocabulary terms	Cosine similarity, IC-based semantic similarity, ranked relevance	(Vandenhende et al., 8 Dec 2025, Painter et al., 26 Mar 2025, Lei et al., 2018)

This variety is substantive rather than terminological. Some indices are symmetric, some are asymmetric; some are thresholded for binary decisions, others are used for ranking or clustering; some operate on raw observations, others on embeddings, ontologies, or conditional densities. A plausible implication is that “Medical Similarity Index” is best understood as a task family rather than a single metric class.

2. Semantic similarity in medical language

In medical NLP, similarity is repeatedly shown to be domain-dependent. For online health questions, a transformer-based classifier can be interpreted as an index over question pairs, with the model output $p(y=1\mid q_1,q_2)$ serving as the probability that two questions are medically equivalent. A central result is that intermediate pre-training on in-domain medical question–answer matching yields markedly better final question similarity performance than either Quora Question Pairs pre-training or no intermediate pre-training: XLNet with medical QA pre-training reaches $82.6\%\pm0.8\%$ accuracy, whereas other approaches remain below $78.7\%$ on the same task (McCreery et al., 2019). This directly contradicts the common assumption that generic paraphrase modeling transfers unchanged to clinical or consumer-health semantics.

Clinical sentence similarity has been formalized more explicitly as an expert-defined ordinal index. MedSTS provides 174,629 sentence pairs from Mayo Clinic notes and an annotated subset of 1,068 pairs scored on a $y=1$0–$y=1$1 scale by two medical experts; weighted Cohen’s $y=1$2 is $y=1$3, which the paper describes as substantial agreement (Wang et al., 2018). The same study shows that simple surface measures correlate only moderately with expert clinical judgments, with Pearson correlations on MedSTS_ann of $y=1$4 for Ratcliff/Obershelp, $y=1$5 for cosine similarity, $y=1$6 for normalized Levenshtein, and $y=1$7 for their ensemble (Wang et al., 2018). This indicates that semantic equivalence in clinical text depends on concept-level content, negation, temporality, and domain-specific synonymy rather than on lexical overlap alone.

Radiology report evaluation pushes the same argument further. For chest X-ray reports, MCSE first extracts enriched clinical entities, including negations and descriptors, and then combines exact matches with cosine similarity between unmatched entities in a biomedical embedding space. The final score

$y=1$8

is normalized to $y=1$9 (Picha et al., 2024). In validation, report pairs with similar CheXpert label patterns lie above roughly $0$0, whereas opposite-label pairs tend to fall below that boundary. On report generation outputs, MCSE scores of $0$1 for R2Gen and $0$2 for CXR-RePaiR coexist with BLEU-2 scores of $0$3 and $0$4, respectively, illustrating that clinically meaningful semantic similarity can remain moderate to high even when n-gram overlap is low (Picha et al., 2024).

Large-scale term similarity benchmarks make the same point at the terminology level. Automatically derived SNOMED datasets contain about $0$5 multi-word terms, and a doctor annotation study over 1,200 sampled pairs reports Krippendorff’s $0$6. The strongest word-embedding results come from MeSH-enhanced FastText models, while contextual embeddings perform poorly on the hardest lexically deceptive splits; the paper concludes that current embeddings remain limited in their ability to adequately encode medical terms (Schulz et al., 2020). Taken together, these studies suggest that any language-oriented Medical Similarity Index must be explicitly domain-relevant in both representation and supervision.

3. Similarity for retrieval, ranking, and longitudinal patients

A second major use of medical similarity indices is retrieval and ranking. In seed-driven document ranking for systematic reviews, Mirror Matching defines similarity between a seed abstract $0$7 and a candidate abstract $0$8 as

$0$9

where both asymmetric components are position-aware averages of maximal embedding-based term matches (Lee et al., 2021). The method is motivated by recurring medical abstract structure—Background, Methods, Results, Conclusion—and by the positional distribution of PICO elements. On the CLEF 2019 TAR dataset, Mirror Matching attains an average precision of $5$0, improving over WMD at $5$1, BM25 at $5$2, QL at $5$3, SDR-BOC at $5$4, and TF-IDF at $5$5 (Lee et al., 2021). Here, the similarity index is less a semantic equivalence score than a retrieval-oriented relevance function.

Patient similarity introduces longitudinality and heterogeneity. A broad review defines patient similarity as a function $5$6 over clinical and biological representations and emphasizes data integration, similarity measurement, and neighborhood identification as the core stages of a similarity study (Dai et al., 2020). Static measures such as Euclidean distance, Mahalanobis distance, cosine similarity, and Jaccard similarity remain important, but the review also highlights supervised metric learning and neural representation learning as central directions (Dai et al., 2020).

Concrete longitudinal formulations show why. One deep EHR architecture represents each patient as a temporally ordered matrix of visit embeddings and computes similarity either through matrix-based dependence measures such as RV and dCor or through a supervised CNN whose patient-level similarity is a bilinear form,

$5$7

On four chronic disease cohorts, the CNN-based method reaches $5$8, $5$9, and $0$0 on the easier dataset, and still improves substantially on the harder dataset where key cohort identifiers are removed (Zhu et al., 2019). A complementary longitudinal approach uses subsequence dynamic time warping so that the shorter patient trajectory is matched to the most relevant subsequence of the longer one rather than globally aligned from first to last visit. In risk stratification for progression from MCI to probable Alzheimer’s disease, subsequence matching reaches an AUROC of $0$1, improving on snapshot data at $0$2 and global alignment at $0$3 (Goyal et al., 2018). These results indicate that, for patient-level indices, temporal misalignment and stage heterogeneity are not secondary nuisances but part of the definition of similarity itself.

4. Morphological and geometric similarity in signals and images

In medical physics and biomedical signal analysis, Medical Similarity Indices often compare shapes rather than semantics. For radiotherapy beam symmetry, the Correlative Symmetric Index compares the mirrored left beam profile with the right profile by cross-correlation,

$0$4

with zero lag used as the meaningful symmetry location, and a normalized form behaving like a correlation coefficient in $0$5 (Bermudez et al., 28 Mar 2025). In noisy non-symmetric simulations, CSI changes by only $0$6, versus $0$7 for area-based symmetry, $0$8 for PDQ, and $0$9 for SSIM; in measured symmetric clinical profiles, all methods are above $5$0, with normalized CSI around $5$1–$5$2 (Bermudez et al., 28 Mar 2025). The paper therefore frames CSI as a noise-robust global similarity/symmetry measure.

For spinal cord injury assessment, a generalized adaptive signed correlation index extends the conventional trichotomized ASCI to $5$3 levels and defines

$5$4

The index remains in $5$5, with $5$6 indicating perfect morphological match and $5$7 maximal mismatch, but the larger $5$8 yields finer resolution of injury severity than the original three-level formulation (Olenko et al., 2016). This is an explicitly morphology-centered similarity index: the closer an injured SEP waveform remains to the healthy reference, the less severe the injury is taken to be.

Multimodal image registration introduces yet another geometric form. Assuming a local functional dependence $5$9 between moving and fixed intensities, the Hessian-based similarity metric derives the relation

$\kappa=\int \sqrt{f_D(y)}\sqrt{f_{\bar D}(y)}\,dy,$0

and then measures the minimized deviation from that relation via a normalized error $\kappa=\int \sqrt{f_D(y)}\sqrt{f_{\bar D}(y)}\,dy,$1, with similarity $\kappa=\int \sqrt{f_D(y)}\sqrt{f_{\bar D}(y)}\,dy,$2 (Eskandari et al., 2023). In MRI–ultrasound registration for image-guided neurosurgery, integrating this metric into an affine framework yields a mean target registration error of $\kappa=\int \sqrt{f_D(y)}\sqrt{f_{\bar D}(y)}\,dy,$3 mm versus $\kappa=\int \sqrt{f_D(y)}\sqrt{f_{\bar D}(y)}\,dy,$4 mm for gradient orientation alignment, with a paired $\kappa=\int \sqrt{f_D(y)}\sqrt{f_{\bar D}(y)}\,dy,$5-test $\kappa=\int \sqrt{f_D(y)}\sqrt{f_{\bar D}(y)}\,dy,$6-value of $\kappa=\int \sqrt{f_D(y)}\sqrt{f_{\bar D}(y)}\,dy,$7 (Eskandari et al., 2023). The metric is thus a second-order structural similarity index rather than an intensity-overlap measure.

The explicitly named Medical Similarity Index for segmentation is boundary-based and clinically tunable. It computes a bidirectional local distance for each test-contour point, converts that signed distance into a point score through a Gaussian weight function, and averages over all test points: $\kappa=\int \sqrt{f_D(y)}\sqrt{f_{\bar D}(y)}\,dy,$8 Separate inside and outside penalty levels, $\kappa=\int \sqrt{f_D(y)}\sqrt{f_{\bar D}(y)}\,dy,$9 and $p(y=1\mid q_1,q_2)$0, allow one to penalize under-segmentation and over-segmentation differently (Fazekas et al., 13 Aug 2025). The prostate example is deliberately clinical: Dice $p(y=1\mid q_1,q_2)$1 and Jaccard $p(y=1\mid q_1,q_2)$2 can coexist with MSI $p(y=1\mid q_1,q_2)$3 when $p(y=1\mid q_1,q_2)$4, because a small outward deviation toward the bladder is clinically unacceptable even if the overlap remains high (Fazekas et al., 13 Aug 2025). This directly challenges the widespread assumption that overlap coefficients alone suffice for medical contour evaluation.

5. Ontology-, vocabulary-, and safety-centered indices

Several works define medical similarity through controlled vocabularies, knowledge graphs, or curated safety terminologies. In pharmacovigilance, SafeTerm embeds MedDRA Preferred Terms and queries into a shared vector space and ranks PTs by cosine similarity. Thresholds on this score define operational relatedness: at $p(y=1\mid q_1,q_2)$5, mean precision is $p(y=1\mid q_1,q_2)$6, recall $p(y=1\mid q_1,q_2)$7, and $p(y=1\mid q_1,q_2)$8; an automatic Knee-based threshold around $p(y=1\mid q_1,q_2)$9 favors recall, yielding precision $82.6\%\pm0.8\%$0 and recall $82.6\%\pm0.8\%$1 (Vandenhende et al., 8 Dec 2025). Narrow-term retrieval is shifted to higher optimal thresholds by about $82.6\%\pm0.8\%$2, which indicates that the same continuous similarity scale can support both broad signal detection and narrow, high-specificity retrieval (Vandenhende et al., 8 Dec 2025).

Ontology-based MedDRA clustering makes the same point with symbolic rather than embedding-driven similarity. Using UMLS-linked MedDRA, SNOMED CT, and MeSH structure, the study evaluates six semantic similarity measures, including path-based LCH and WUPALMER and intrinsic IC-based RESNIK, LIN, INTRINSIC-LCH, and SOKAL. The best clustering accuracy comes from IC-based measures, with maximum $82.6\%\pm0.8\%$3 around $82.6\%\pm0.8\%$4 for INTRINSIC-LIN and SOKAL, compared with $82.6\%\pm0.8\%$5 for WUPALMER and $82.6\%\pm0.8\%$6 for LCH (Painter et al., 26 Mar 2025). This suggests that, in drug safety, shared information content at the lowest common ancestor is more useful than pure path length when the task is grouping PTs around medically meaningful centroids.

Drug-level similarity in MedSim is broader still. It fuses medicine-specific features—side effects, targets, mechanisms of action, physiologic effects—with hierarchy embeddings, BM25-based textual similarity, and word embeddings, then learns the final similarity score by random forest regression. On 528 antibiotic pairs scored by doctors, MedSim reaches Pearson $82.6\%\pm0.8\%$7 and Spearman $82.6\%\pm0.8\%$8, substantially above GADES, Resnik, Wpath, and other baselines (Lei et al., 2018). The case study proposes that antibiotics with similarity score $82.6\%\pm0.8\%$9 may be considered substitutable under normal circumstances, illustrating how a continuous similarity index becomes an operational decision rule (Lei et al., 2018).

Knowledge-graph similarity also appears in medical image retrieval evaluation. There, images are represented by sets of UMLS CUIs extracted from captions, and exact IoU is extended to nearest-neighbor IoU: $78.7\%$0 where $78.7\%$1 contains non-identical concepts whose graph distance is within a threshold $78.7\%$2 (Wei et al., 16 Jun 2025). Embedded into an NDCG-style retrieval score, this yields a label-free semantically aware evaluation for CBIR; with modality+organ precision at $78.7\%$3, nn-IoU reaches $78.7\%$4 versus $78.7\%$5 for plain IoU (Wei et al., 16 Jun 2025). Here, the similarity index measures graded conceptual proximity rather than literal label identity.

6. Distributional interpretation, thresholds, and general limits

Some of the most abstract similarity indices compare not two objects but two medically relevant distributions. In diagnostic testing, the affinity

$78.7\%$6

is the Hellinger affinity between diseased and non-diseased biomarker distributions, with $78.7\%$7 indicating perfect separation and $78.7\%$8 indicating identical distributions (Carvalho et al., 2017). The covariate-specific version $78.7\%$9 extends this idea conditionally. In a prostate cancer study, total PSA yields $y=1$00 and AUC $y=1$01, whereas the free-to-total PSA ratio yields $y=1$02 and AUC $y=1$03, so the affinity index captures poorer discrimination as greater distributional similarity (Carvalho et al., 2017).

Dataset-level diversity assessment uses a related but higher-order construction. SDICE first computes similarity scores between image embeddings from a domain-specific contrastive encoder, then compares the induced similarity-score distributions for real and synthetic data. After exponential normalization relative to a worst-case “transformed image” baseline, it reports intra-class and inter-class diversity indices $y=1$04 (Alam et al., 2024). On MIMIC-CXR, $y=1$05 and $y=1$06, whereas on ImageNet the corresponding values are about $y=1$07 and $y=1$08 (Alam et al., 2024). The interpretation is direct: synthetic chest X-rays preserve class separation better than within-class diversity.

Taken together, these works suggest several limits on any universal notion of a Medical Similarity Index. First, similarity is object-dependent: semantic equivalence of clinical sentences, covariance of beam halves, overlap of diagnostic densities, and contour safety in radiotherapy are not reducible to one shared geometry. Second, threshold choice is intrinsic rather than ancillary. SafeTerm changes behavior substantially between $y=1$09 and $y=1$10 (Vandenhende et al., 8 Dec 2025); MCSE uses a practical boundary around $y=1$11 for report similarity (Picha et al., 2024); segmentation MSI changes sharply with $y=1$12 and $y=1$13 (Fazekas et al., 13 Aug 2025). Third, annotation and ontology themselves introduce uncertainty: expert agreement on medical question equivalence is below perfect, MedSTS annotators report difficulty with fine distinctions, and ontology-based safety clustering does not fully coincide with expert review (McCreery et al., 2019, Wang et al., 2018, Painter et al., 26 Mar 2025).

A plausible overall implication is that a Medical Similarity Index is best treated as a domain-calibrated scoring function whose validity depends on three coupled choices: the medical object being compared, the representation used to encode it, and the clinical decision that the score is meant to support.