medact(feat): cdb utils for merging and navigation using pt2ch relations

medcat-v2/medcat/utils/cdb_utils.py

@@ -0,0 +1,292 @@
+from collections import defaultdict
+import logging
+import numpy as np
+
+from copy import deepcopy
+from medcat.cdb import CDB
+
+logger = logging.getLogger(__name__)  # separate logger from the package-level one
+
+
+def merge_cdb(cdb1: CDB,
+              cdb2: CDB,
+              overwrite_training: int = 0,
+              full_build: bool = False) -> CDB:
+    """Merge two CDB's together to produce a new, single CDB. The contents of
+    inputs CDBs will not be changed.
+    `addl_info` can not be perfectly merged, and will prioritise cdb1. see `full_build`
+
+    Args:
+        cdb1 (CDB):
+            The first medcat cdb to merge. In cases where merging isn't suitable
+            isn't ideal (such as cui2preferred_name), this cdb values will be
+            prioritised over cdb2.
+        cdb2 (CDB):
+            The second medcat cdb to merge.
+        overwrite_training (int):
+            Choose to prioritise a CDB's context vectors values over merging gracefully.
+            0 - no prio, 1 - CDB1, 2 - CDB2
+        full_build (bool):
+            Add additional information from "addl_info" dicts "cui2ontologies" and
+            "cui2description"
+
+    Returns:
+        CDB: The merged CDB.
+    """
+    config = deepcopy(cdb1.config)
+    cdb = CDB(config)
+
+    # Copy CDB 1 - as all settings from CDB 1 will be carried over
+    cdb.cui2info = deepcopy(cdb1.cui2info)
+    cdb.name2info = deepcopy(cdb1.name2info)
+    cdb.type_id2info = deepcopy(cdb1.type_id2info)
+    cdb.token_counts = deepcopy(cdb1.token_counts)
+    cdb._subnames = deepcopy(cdb1._subnames)
+    if full_build:
+        cdb.addl_info = deepcopy(cdb1.addl_info)
+
+    # Merge concepts from cdb2 into the merged CDB
+    for cui, cui_info2 in cdb2.cui2info.items():
+        # Get name status from cdb2
+        name_status = 'A'  # default status
+        for name in cui_info2['names']:
+            if name in cdb2.name2info:
+                name_info = cdb2.name2info[name]
+                if cui in name_info['per_cui_status']:
+                    name_status = name_info['per_cui_status'][cui]
+                    break
+
+        # Prepare names dict for _add_concept
+        names = {}
+        for name in cui_info2['names']:
+            # Create a simple NameDescriptor-like structure
+            names[name] = type('NameDescriptor', (), {
+                'snames': cui_info2['subnames'],
+                'is_upper': cdb2.name2info.get(name, {}).get('is_upper', False),
+                'tokens': set(),  # We don't have token info in the new structure
+                'raw_name': name
+            })()
+
+        # Get ontologies and description for full_build
+        ontologies = set()
+        description = cui_info2.get('description', '')
+        to_build = full_build and (
+            cui_info2.get('original_names') is not None or
+            cui_info2.get('description') is not None
+        )
+
+        if to_build and cui_info2.get('in_other_ontology'):
+            ontologies.update(cui_info2['in_other_ontology'])
+
+        cdb._add_concept(
+            cui=cui, names=names, ontologies=ontologies, name_status=name_status,
+            type_ids=cui_info2['type_ids'], description=description,
+            full_build=to_build
+        )
+
+        # Copy training data from cdb2 for concepts that don't exist in cdb1
+        if cui not in cdb1.cui2info:
+            cui_info_merged = cdb.cui2info[cui]
+            cui_info_merged['count_train'] = cui_info2['count_train']
+            cui_info_merged['context_vectors'] = deepcopy(cui_info2['context_vectors'])
+            cui_info_merged['average_confidence'] = cui_info2['average_confidence']
+            if cui_info2.get('tags'):
+                cui_info_merged['tags'] = deepcopy(cui_info2['tags'])
+
+        # Handle merging of training data for concepts that exist in both CDBs
+        if cui in cdb1.cui2info:
+            cui_info1 = cdb1.cui2info[cui]
+            cui_info_merged = cdb.cui2info[cui]
+
+            # Merge count_train
+            if (cui_info1['count_train'] > 0 or cui_info2['count_train'] > 0) and not (
+                overwrite_training == 1 and cui_info1['count_train'] > 0
+            ):
+                if overwrite_training == 2 and cui_info2['count_train'] > 0:
+                    cui_info_merged['count_train'] = cui_info2['count_train']
+                else:
+                    cui_info_merged['count_train'] = (
+                        cui_info1['count_train'] + cui_info2['count_train']
+                    )
+
+            # Merge context vectors
+            if (cui_info1['context_vectors'] is not None and
+                not (overwrite_training == 1 and
+                     cui_info1['context_vectors'] is not None)):
+
+                if (overwrite_training == 2 and
+                    cui_info2['context_vectors'] is not None):
+                    cui_info_merged['context_vectors'] = deepcopy(
+                        cui_info2['context_vectors']
+                    )
+                else:
+                    # Merge context vectors with weighted average
+                    if cui_info_merged['context_vectors'] is None:
+                        cui_info_merged['context_vectors'] = {}
+
+                    # Get all context types from both CDBs
+                    contexts = set()
+                    if cui_info1['context_vectors']:
+                        contexts.update(cui_info1['context_vectors'].keys())
+                    if cui_info2['context_vectors']:
+                        contexts.update(cui_info2['context_vectors'].keys())
+
+                    # Calculate weights
+                    if overwrite_training == 2:
+                        weights = [0, 1]
+                    else:
+                        norm = cui_info_merged['count_train']
+                        if norm > 0:
+                            weights = [
+                                np.divide(cui_info1['count_train'], norm),
+                                np.divide(cui_info2['count_train'], norm)
+                            ]
+                        else:
+                            weights = [0.5, 0.5]  # equal weights if no training
+
+                    # Merge each context vector
+                    for context_type in contexts:
+                        if cui_info1['context_vectors']:
+                            vec1 = cui_info1['context_vectors'].get(
+                                context_type, np.zeros(300)
+                            )
+                        else:
+                            vec1 = np.zeros(300)
+
+                        if cui_info2['context_vectors']:
+                            vec2 = cui_info2['context_vectors'].get(
+                                context_type, np.zeros(300)
+                            )
+                        else:
+                            vec2 = np.zeros(300)
+                        cui_info_merged['context_vectors'][context_type] = (
+                            weights[0] * vec1 + weights[1] * vec2
+                        )
+
+            # Merge tags
+            if cui_info1.get('tags') and cui_info2.get('tags'):
+                if cui_info_merged['tags'] is None:
+                    cui_info_merged['tags'] = []
+                cui_info_merged['tags'].extend(cui_info2['tags'])
+
+            # Merge type_ids (already handled by _add_concept, but ensure union)
+            cui_info_merged['type_ids'].update(cui_info2['type_ids'])
+
+    # Merge name training counts
+    if overwrite_training != 1:
+        for name, name_info2 in cdb2.name2info.items():
+            if name in cdb1.name2info and overwrite_training == 0:
+                # Merge training counts for names that exist in both CDBs
+                name_info1 = cdb1.name2info[name]
+                name_info_merged = cdb.name2info[name]
+                name_info_merged['count_train'] = (
+                    name_info1['count_train'] + name_info2['count_train']
+                )
+            else:
+                # Copy name info from cdb2 if it doesn't exist in cdb1
+                if name not in cdb.name2info:
+                    cdb.name2info[name] = deepcopy(name_info2)
+
+    # Merge token counts
+    if overwrite_training != 1:
+        for token, count in cdb2.token_counts.items():
+            if token in cdb.token_counts and overwrite_training == 0:
+                cdb.token_counts[token] += count
+            else:
+                cdb.token_counts[token] = count
+
+    return cdb
+
+
+def _dedupe_preserve_order(items: list[str]) -> list[str]:
+    seen = set()
+    deduped_list = []
+    for item in items:
+        if item not in seen:
+            seen.add(item)
+            deduped_list.append(item)
+    return deduped_list
+
+
+def get_all_ch(parent_cui: str, cdb):
+    """Get all the children of a given parent CUI. Preserves the order of the parent
+
+    Args:
+        parent_cui (str): The parent CUI
+        cdb (CDB): The CDB object
+
+    Returns:
+        list: The children of the parent CUI
+    """
+    all_ch = [parent_cui]
+    for cui in cdb.addl_info.get('pt2ch', {}).get(parent_cui, []):
+        cui_chs = get_all_ch(cui, cdb)
+        all_ch += cui_chs
+    return _dedupe_preserve_order(all_ch)
+
+
+def ch2pt_from_pt2ch(cdb: CDB, pt2ch_key: str = 'pt2ch'):
+    """Get the child to parent info from the pt2ch map in the CDB
+
+    Args:
+        cdb (CDB): The CDB object with addl_info['pt2ch']
+        pt2ch_key (str, optional): The key in the addl_info dict to get the pt2ch map
+            from.
+            Defaults to 'pt2ch'.
+    Returns:
+        dict: The child to parent info
+    """
+    ch2pt = defaultdict(list)
+    for k, vals in cdb.addl_info[pt2ch_key].items():
+        for v in vals:
+            ch2pt[v].append(k)
+    return ch2pt
+
+
+def snomed_ct_concept_path(
+    cui: str, cdb: CDB, parent_node='138875005'
+) -> dict[str, list[dict]]:
+    """Get the concept path for a given CUI to a parent node
+
+    Args:
+        cui (str): The CUI of the concept to get the path for
+        cdb (CDB): The CDB object
+        parent_node (str, optional): The top level parent node.
+            Defaults to '138875005' the root SNOMED CT code.
+
+    Returns:
+        dict: The concept path and links
+    """
+    try:
+        def find_parents(cui, cuis2nodes, child_node=None):
+            parents = list(cdb.addl_info.get('ch2pt', {}).get(cui, []))
+            all_links = []
+            if cui not in cuis2nodes:
+                # Get preferred name from the new CDB structure
+                preferred_name = cdb.cui2info.get(cui, {}).get('preferred_name', cui)
+                curr_node = {'cui': cui, 'pretty_name': preferred_name}
+                if child_node:
+                    curr_node['children'] = [child_node]
+                cuis2nodes[cui] = curr_node
+                if len(parents) > 0:
+                    all_links += find_parents(
+                        parents[0], cuis2nodes, child_node=curr_node
+                    )
+                    for p in parents[1:]:
+                        links = find_parents(p, cuis2nodes)
+                        all_links += [{'parent': p, 'child': cui}] + links
+            else:
+                if child_node:
+                    if 'children' not in cuis2nodes[cui]:
+                        cuis2nodes[cui]['children'] = []
+                    cuis2nodes[cui]['children'].append(child_node)
+            return all_links
+        cuis2nodes = dict()
+        all_links = find_parents(cui, cuis2nodes)
+        return {
+            'node_path': cuis2nodes[parent_node],
+            'links': all_links
+        }
+    except KeyError as e:
+        logger.warning(f'Cannot find path concept path:{e}')
+        return []