RareDx / src /rag.py
sarra2002's picture
Démo RareDx
613af80 verified
Raw History Blame Contribute Delete
2.03 kB
"""RAG branch: hierarchical, information-weighted search over the disease profiles.
For each symptom of the patient, a disease receives the information content (IC) of the
most specific common ancestor; scores are summed and slightly penalised for very long
profiles. This is a variant of the Phenomizer / Resnik similarity.
"""
from __future__ import annotations
from collections import defaultdict
import numpy as np
from .knowledge_base import KnowledgeBase
class HierarchicalRAG:
def __init__(self, kb: KnowledgeBase, alpha: float = 0.25):
self.kb = kb
self.alpha = alpha
self.ids = [d["code"] for d in kb.diseases]
extended = [kb.expand([s["hpo"] for s in d["symptoms"]]) for d in kb.diseases]
index: dict[str, list[int]] = defaultdict(list)
for i, terms in enumerate(extended):
for t in terms:
index[t].append(i)
self.index = {t: np.array(v) for t, v in index.items()}
self.n = len(self.ids)
self.ic = {t: -np.log(len(v) / self.n) for t, v in self.index.items()}
self.size = np.array([len(e) for e in extended])
def scores(self, codes: list[str]) -> np.ndarray:
total = np.zeros(self.n)
for h in codes:
best = np.zeros(self.n)
for a in self.kb.ancestors(h):
if a in self.index:
idx = self.index[a]
best[idx] = np.maximum(best[idx], self.ic[a])
total += best
return total / np.log(self.size + np.e) ** self.alpha
def search(self, codes: list[str], k: int = 20) -> list[str]:
"""HPO codes -> the k best diseases (distinct groups), most likely first."""
scores = self.scores(codes)
result, seen = [], set()
for i in np.argsort(-scores):
g = self.kb.group(self.ids[i])
if g not in seen:
seen.add(g)
result.append(g)
if len(result) == k:
break
return result