Rare-Disease Diagnosis on the ZebraMap Multimodal Case Report Dataset: A Hybrid Pipeline with Grounded Explainability

. 2026 Jun 4;26(11):3582. doi: 10.3390/s26113582

Algorithm 1 Hybrid Candidate Generation and Reciprocal-Rank Fusion

Input: case query $q (c)$ , disease catalog $D$ , per-source depth $K_{s}$ , final depth $K_{f}$ .
Output: hybrid candidate list $π_{c}^{(0)}$ .
1. Score all diseases in $D$ with BM25 over disease-profile text and retain the top $K_{s}$ results.
2. Transform $q (c)$ with the fitted TF–IDF vectorizer, predict class probabilities with the class-balanced linear classifier, and retain the top $K_{s}$ diseases.
3. Encode $q (c)$ and all disease profiles with MedCPT, compute cosine similarity, and retain the top $K_{s}$ diseases.
4. Merge the three ranked lists with RRF using source weights ${1.5, 0.75, 1.0}$ and constant $K = 60$ .
5. Sort diseases by $s_{RRF}$ and retain the top $K_{f}$ candidates.