import os import re import json import joblib import numpy as np from crewai import Agent, Task, Crew, LLM from crewai.tools import tool from firebaseLoader import pull_audio, pull_lifestyle from webscraper import * from typing import Union import shap import xgboost as xgb from xgboost import DMatrix from math import exp # ================================================================ # Loading in the prediction models # ================================================================ # Load audio model & metadata audio_model = xgb.Booster() audio_model.load_model('models/modelAudiov8_ML.json') metadata = joblib.load('models/modelAudiov8_metadata.pk1') audio_feature_names = metadata['feature_names'] audio_mmse_index = metadata['mmse_index'] # Load lifestyle model & metadata lifestyle_model = xgb.Booster() lifestyle_model.load_model('models/modelLifestyleV3_ML.json') metadata = joblib.load('models/modelLifestyleV3_metadata.pk1') lifestyle_feature_names = metadata['feature_names'] # ================================================================ # Loading in the SHAP explainers # ================================================================ # Load model explainers audio_explainer = shap.TreeExplainer(audio_model) lifestyle_explainer = shap.TreeExplainer(lifestyle_model) placeholder_lifestyle_ID = "umWSG13rLSfHqqJlq1WluPzjLB93" # Load 2 different llms for different purposes. llm_predict = LLM(model="huggingface/openai/gpt-oss-120b") llm_news = LLM(model="huggingface/Qwen/Qwen3-4B-Thinking-2507") # ================================================================ # Defining the tools for the agents to make use of # ================================================================ # Prediction tool using the audio model @tool("calculate_dementia_risk_audio") def calculate_dementia_risk_audio(docID: str) -> Union[str, dict]: """ Calculates dementia risk for a specific patient using audio and transcript data from Firebase. Returns a structured dict. """ try: # Handle JSON-wrapped input (from CrewAI) if docID.strip().startswith('{'): try: parsed = json.loads(docID) if isinstance(parsed, dict): if "Audio" in parsed and isinstance(parsed["Audio"], dict): docID = parsed["Audio"].get("docID", docID) elif "docID" in parsed: docID = parsed["docID"] except json.JSONDecodeError: pass # Pulls data from firebase using the patientID (docID) audio_data = pull_audio(docID) # Error handling for missing patient entry if audio_data is None: error_msg = f"Error: Could not retrieve or process audio data for patient '{docID}'" return {"status": "error", "message": error_msg} # Extract values age = audio_data.get("age") gender_val = audio_data.get("gender") mmse = audio_data.get("mmse") # This feature was defined as optional during the audio model training combined_features = audio_data.get("full_features") if combined_features is None: error_msg = f"Error: Feature extraction failed for patient '{docID}'" return {"status": "error", "message": error_msg} # Convert the numpy array to DMatrix for compatability audio_array = np.array(combined_features).reshape(1, -1) dmat = DMatrix(audio_array) # Predict using the patient data, returns a 1 or 0 prediction = audio_model.predict(dmat)[0] # Getting the confidence (probability) of the prediction try: # If model supports predict_proba (e.g., scikit-learn wrapper), use it proba = audio_model.predict_proba(dmat)[0] confidence = float(max(proba)) except AttributeError: # Fallback, use raw prediction scores (if binary classification) raw_score = audio_model.predict(dmat, output_margin=True)[0] # Convert to probability via sigmoid (for binary classification) proba_pos = 1 / (1 + exp(-raw_score)) confidence = proba_pos if prediction == 1 else (1 - proba_pos) # Setting the risk level risk_level = "High" if prediction == 1 else "Low" gender_str = "Male" if gender_val == 1 else "Female" # Structured output for easy parsing result_dict = { "risk_level": risk_level, "prediction": int(prediction), "confidence": confidence, "age": age, "gender": gender_str, "docID": docID } return {"status": "success", **result_dict} # Handles failure to predict except Exception as e: error_msg = f"Error calculating audio-based risk: {str(e)}" return {"status": "error", "message": error_msg} # SHAP explainer tool for audio prediction results @tool("explain_dementia_risk_audio") def explain_dementia_risk_audio(docID: str) -> str: """Provides SHAP-based explanation for audio risk prediction.""" try: # Pulls data from firebase using the patientID (docID) audio_data = pull_audio(docID) # Handles missing patient entry if audio_data is None: return f"Error: No data for {docID}" # Get the patients data features = audio_data["full_features"] feature_names = audio_feature_names # SHAP values for this instance shap_values = audio_explainer.shap_values(features.reshape(1, -1)) # shap_values is array of shape (1, n_features) for class 1 (Dementia) # Get top 5 positive contributors (pushing toward Dementia = class 1) shap_vals = shap_values[0] # for class 1 top_indices = np.argsort(-shap_vals)[:5] # descending by impact # list of shap explanations explanation_lines = ["Top factors increasing dementia risk (audio analysis):"] for i in top_indices: feat = feature_names[i] val = features[i] impact = shap_vals[i] if abs(impact) < 1e-4: continue direction = "↑" if impact > 0 else "↓" explanation_lines.append(f" • {feat} = {val:.3f} → {direction} risk ({impact:+.3f})") return "\n".join(explanation_lines) # Error handling except Exception as e: return f"SHAP explanation failed: {str(e)}" @tool("calculate_dementia_risk_lifestyle") def calculate_dementia_risk_lifestyle(docID: str) -> Union[str, dict]: """ Calculates dementia risk for a specific patient using lifestyle data. Returns a structured dict """ try: # Handle JSON-wrapped input (from CrewAI) if docID.strip().startswith('{'): try: parsed = json.loads(docID) if isinstance(parsed, dict): if "Lifestyle" in parsed and isinstance(parsed["Lifestyle"], dict): docID = parsed["Lifestyle"].get("docID", docID) elif "docID" in parsed: docID = parsed["docID"] except json.JSONDecodeError: pass # Pulls data from firebase using the patientID (docID) lifestyle_data = pull_lifestyle(docID) if not lifestyle_data: error = f"Error: No lifestyle data found for document ID '{docID}'" return {"status": "error", "message": error} # feature names in firebase dont match feature names in lifestyle model, so those names need to be changed FEATURE_MAPPING = { 'Age': 'age', 'Gender': 'gender', 'Ethnicity': 'ethnicity', 'EducationLevel': 'education_level', 'BMI': 'bmi', 'Smoking': 'smoking', 'AlcoholConsumption': 'alcohol_consumption', 'PhysicalActivity': 'physical_activity', 'DietQuality': 'diet_quality', 'SleepQuality': 'sleep_quality', 'FamilyHistoryAlzheimers': 'family_history_dementia', 'CardiovascularDisease': 'cardiovascular_disease', 'Diabetes': 'diabetes', 'Depression': 'depression', 'HeadInjury': 'head_injury', 'Hypertension': 'hypertension' } features = [] for feat_name in lifestyle_feature_names: firestore_key = FEATURE_MAPPING.get(feat_name) if firestore_key is None: error = f"Error: No mapping defined for feature '{feat_name}'" return {"status": "error", "message": error} val = lifestyle_data.get(firestore_key) if val is None: error = f"Error: Missing required lifestyle feature '{feat_name}' (Firestore key: '{firestore_key}')" return {"status": "error", "message": error} features.append(val) # Convert to DMatrix life_array = np.array(features).reshape(1, -1) dmat = DMatrix(life_array) # Predict prediction = lifestyle_model.predict(dmat)[0] # For confidence try: # If model supports predict_proba (e.g., scikit-learn wrapper), use it proba = lifestyle_model.predict_proba(dmat)[0] confidence = float(max(proba)) except AttributeError: # Fallback: use raw prediction scores (if binary classification) raw_score = lifestyle_model.predict(dmat, output_margin=True)[0] # Convert to probability via sigmoid (for binary classification) proba_pos = 1 / (1 + exp(-raw_score)) confidence = proba_pos if prediction == 1 else (1 - proba_pos) gender_val = lifestyle_data.get("gender") age = lifestyle_data.get("age") # Setting the risk level risk_level = "High" if prediction == 1 else "Low" gender_str = "Male" if gender_val == 1 else "Female" # Structured output for easy parsing result_dict = { "risk_level": risk_level, "prediction": int(prediction), "confidence": confidence, "age": age, "gender": gender_str, "docID": docID } return {"status": "success", **result_dict} # Handles failure to predict except Exception as e: error_msg = f"Error calculating risk: {str(e)}" return {"status": "error", "message": error_msg} @tool("explain_dementia_risk_lifestyle") def explain_dementia_risk_lifestyle(docID: str) -> str: """Provides SHAP-based explanation for lifestyle risk prediction.""" try: # Pulls data from firebase using the patientID (docID) lifestyle_data = pull_lifestyle(docID) # Handles missing patient entry if not lifestyle_data: return f"Error: No lifestyle data for {docID}" # Remapping the feature names to work with lifestyle model FEATURE_MAPPING = { 'Age': 'age', 'Gender': 'gender', 'Ethnicity': 'ethnicity', 'EducationLevel': 'education_level', 'BMI': 'bmi', 'Smoking': 'smoking', 'AlcoholConsumption': 'alcohol_consumption', 'PhysicalActivity': 'physical_activity', 'DietQuality': 'diet_quality', 'SleepQuality': 'sleep_quality', 'FamilyHistoryAlzheimers': 'family_history_dementia', 'CardiovascularDisease': 'cardiovascular_disease', 'Diabetes': 'diabetes', 'Depression': 'depression', 'HeadInjury': 'head_injury', 'Hypertension': 'hypertension' } features = [] for feat_name in lifestyle_feature_names: val = lifestyle_data.get(FEATURE_MAPPING[feat_name]) if val is None: return f"Missing feature: {feat_name}" features.append(val) X_instance = np.array(features).reshape(1, -1) # Get top 5 positive contributors (pushing toward Dementia = class 1) shap_vals = lifestyle_explainer.shap_values(X_instance)[0] top_indices = np.argsort(-shap_vals)[:5] # descending by impact to prediction result # list of shap explanations explanation_lines = ["Top factors increasing dementia risk (lifestyle analysis):"] for i in top_indices: feat = lifestyle_feature_names[i] val = features[i] impact = shap_vals[i] if abs(impact) < 1e-4: continue # Map categorical to readable (optional but nice) readable_val = val if feat == "Gender": readable_val = "Male" if val == 1 else "Female" elif feat in ["Smoking", "AlcoholConsumption", "FamilyHistoryAlzheimers", "CardiovascularDisease", "Diabetes", "Depression", "HeadInjury", "Hypertension"]: readable_val = "Yes" if val == 1 else "No" elif feat == "EducationLevel": edu_map = {1: "Primary", 2: "Secondary", 3: "Bachelor", 4: "Postgrad"} readable_val = edu_map.get(val, val) direction = "↑" if impact > 0 else "↓" explanation_lines.append(f" • {feat} = {readable_val} → {direction} risk ({impact:+.3f})") return "\n".join(explanation_lines) # Error handling except Exception as e: return f"SHAP explanation failed: {str(e)}" @tool("gather_news_patient") def gather_news_patient(count: int = 5) -> dict: """ Fetches count recent dementia news items for patients, retrieves full text, and generates patient-friendly summaries using the LLM. Returns a json object with the format: {"status": "success", "articles": [{"title", "summary", "url"}, ...]}. """ try: # gets the article data from webscraper raw_articles = get_patient_news_raw(count=count) results = [] for item in raw_articles: url = item["url"] title = item["title"] full_text = scrape_full_text(url) # LM summarization happens using llm_news if not full_text.strip(): summary = "Summary unavailable." else: # Use llm_news — ensure it's in scope prompt = f"""You are a compassionate health educator. Summarize the following article in 2–3 clear, empathetic sentences for patients/families affected by dementia. Focus on: What does this mean? What can they do? Hope & support. Avoid jargon. Article: {full_text[:3000]}""" # Gets the llm_news the create a summary response = llm_news.invoke(prompt) summary = str(response).strip() results.append({ "title": title, "summary": summary, "url": url }) # Structured output for easy parsing return { "status": "success", "articles": results } # Error handling except Exception as e: return { "status": "error", "message": str(e), "articles": [] } @tool("gather_news_doctor") def gather_news_doctor(count: int = 5) -> dict: """ Fetches count recent dementia news items for doctors, retrieves full text, and generates clinical summaries using the LLM. Returns a json object with the format: {"status": "success", "articles": [{"title", "summary", "url"}, ...]}. """ try: # gets the article data from webscraper raw_articles = get_doctor_news_raw(count=count) results = [] for item in raw_articles: url = item["url"] title = item["title"] full_text = scrape_full_text(url) # LM summarization happens using llm_news if not full_text.strip(): summary = "Summary unavailable." else: # Use llm_news — ensure it's in scope prompt = f"""You are a clinical expert. Summarize the following article in 2–3 precise, evidence-based sentences for physicians. Include: key findings, methods (if applicable), clinical relevance. Article: {full_text[:3000]}""" # Gets the llm_news the create a summary response = llm_news.invoke(prompt) summary = str(response).strip() results.append({ "title": title, "summary": summary, "url": url }) # Structured output for easy parsing return { "status": "success", "articles": results } # Error handling except Exception as e: return { "status": "error", "message": str(e), "articles": [] } # Define agents for Alzheimer's detection system risk_calculator_audio = Agent( role="Risk Calculator Audio Agent", goal="Calculate dementia risk for a specific patient using ML model and LLM interpretation", backstory="You combine machine learning predictions with natural language explanations to provide personalized risk assessments.", verbose=True, llm=llm_predict, tools=[calculate_dementia_risk_audio, explain_dementia_risk_audio] ) risk_calculator_lifestyle = Agent( role="Risk Calculator Lifestyle Agent", goal="Calculate dementia risk for a specific patient using ML model and LLM interpretation", backstory="You combine machine learning predictions with natural language explanations to provide personalized risk assessments.", verbose=True, llm=llm_predict, tools=[calculate_dementia_risk_lifestyle, explain_dementia_risk_lifestyle] ) patient_news_agent = Agent( role="Patient News Agent", goal="Gather and summarize dementia/caregiving news for patients and families in plain, supportive language.", backstory="You find real, current articles from trusted sites like caregiving.org, extract full content, and use your medical communication expertise to make it clear and kind for non-experts.", tools=[gather_news_patient], llm=llm_news, verbose=True ) doctor_news_agent = Agent( role="Doctor News Agent", goal="Gather and summarize high-impact dementia research for clinicians and scientists.", backstory="You retrieve peer-reviewed research (e.g., from Europe PMC), extract key findings, and distill them into clinically actionable summaries for time-pressed professionals.", tools=[gather_news_doctor], llm=llm_news, verbose=True ) # ================================================================ # Define tasks for agents # ================================================================ # Audio prediction risk_assessment_task_audio = Task( description=f"Calculate dementia risk for patient", expected_output="Risk level, confidence score, and plain-language explanation.", agent=risk_calculator_audio, ) explanation_task_audio = Task( description="Explain the dementia risk prediction for patient using SHAP", expected_output="Top 5 features increasing risk from audio analysis.", agent=risk_calculator_audio ) # Lifestyle prediction risk_assessment_task_lifestyle = Task( description=f"Calculate dementia risk for patient using the given patient id: {placeholder_lifestyle_ID}", expected_output="Risk level, confidence score, and plain-language explanation.", agent=risk_calculator_lifestyle, ) explanation_task_lifestyle = Task( description=f"Explain the dementia risk prediction for patient {placeholder_lifestyle_ID} using SHAP", expected_output="Top 5 features increasing risk, with impact direction and magnitude.", agent=risk_calculator_lifestyle ) # News gathering get_news_task_patient = Task( description=( "Use the `gather_news_patient` tool to fetch {count} recent dementia/caregiving news articles " "and return the raw JSON result. Do not generate or modify content." "do not add any formatting to the result, such as backticks, only return the raw json dict" "do not modify the given urls or titles" ), expected_output="A JSON object with 'status' and 'articles': list of {title, summary, url}.", agent=patient_news_agent ) get_news_task_doctor = Task( description=( "Use the `gather_news_doctor` tool to fetch {count} recent dementia research articles " "and return the raw JSON result. Do not generate or modify content." "do not add any formatting to the result, such as backticks, only return the raw json dict" "do not modify the given urls or titles" ), expected_output="A JSON object with 'status' and 'articles': list of {title, summary, url}.", agent=doctor_news_agent ) # Human-readable mapping for audio features, as the names are not very understandable. AUDIO_FEATURE_DESCRIPTIONS = { # MFCC means (13 coefficients) "mfcc_mean_0": "Overall speech energy (loudness stability)", "mfcc_mean_1": "Voice pitch stability (fundamental frequency variation)", "mfcc_mean_2": "Vocal tract shape — openness of mouth/jaw", "mfcc_mean_3": "Tongue position — front vs. back articulation", "mfcc_mean_4": "Nasality (airflow through nose during speech)", "mfcc_mean_5": "Spectral tilt — balance of high/low frequencies", "mfcc_mean_6": "Vocal effort / breathiness", "mfcc_mean_7": "Plosive clarity (e.g., 'p', 't', 'k' sounds)", "mfcc_mean_8": "Fricative precision (e.g., 's', 'f', 'sh' sounds)", "mfcc_mean_9": "Vowel space contraction (reduced vowel distinctness)", "mfcc_mean_10": "Speech rate consistency", "mfcc_mean_11": "Pause frequency during speech", "mfcc_mean_12": "Prosodic flatness (monotone speech tendency)", # MFCC stds (variability) "mfcc_std_0": "Loudness variability", "mfcc_std_1": "Pitch variability (intonation range)", "mfcc_std_2": "Vocal tract flexibility", "mfcc_std_3": "Articulatory precision variability", "mfcc_std_4": "Nasal resonance variability", "mfcc_std_5": "Spectral balance variability", "mfcc_std_6": "Breath support consistency", "mfcc_std_7": "Consonant burst strength variation", "mfcc_std_8": "Fricative noise stability", "mfcc_std_9": "Vowel distinctness consistency", "mfcc_std_10": "Speech rhythm stability", "mfcc_std_11": "Pause placement consistency", "mfcc_std_12": "Emotional prosody variation", # Other features "zcr": "Zero-crossing rate — measures speech turbulence/noisiness (e.g., breathiness, hoarseness)", "rms": "Root-mean-square energy — average speech loudness" } # ================================================================ # Helper functions # ================================================================ # Converts the raw explanations to something more understandable for patients/doctors def parse_shap_explanation(shap_text: str, is_lifestyle: bool = True) -> dict: """ Converts raw string SHAP explanation into structured JSON with human-readable feature names. """ lines = shap_text.strip().split("\n") features = [] explanations = [] # Skip header start_idx = 0 for i, line in enumerate(lines): if line.startswith(" •") or "→" in line: start_idx = i break for line in lines[start_idx:]: line = line.strip() if not (line.startswith("•") or line.startswith(" •")): continue line = line.lstrip("• *→↑↓").strip() try: if " → " in line: feature_part, impact_part = line.split(" → ", 1) else: feature_part, impact_part = line, "" if " = " in feature_part: feat_name_raw, feat_val = feature_part.split(" = ", 1) feat_name_raw = feat_name_raw.strip() feat_val = feat_val.strip() else: feat_name_raw = feature_part.strip() feat_val = "unknown" impact_val = "unknown impact" if "(" in impact_part and ")" in impact_part: impact_val = impact_part.split("(")[-1].split(")")[0].strip() # Map to human-readable name if is_lifestyle: feat_name_human = feat_name_raw else: # Audio: map mfcc_mean_1 to "Voice pitch stability", etc. feat_name_human = AUDIO_FEATURE_DESCRIPTIONS.get(feat_name_raw, feat_name_raw).split(" (")[0].split(" —")[0].split(" -")[0].strip() # e.g., from "Voice pitch stability (fundamental frequency variation)" to "Voice pitch stability" features.append(feat_name_human) if is_lifestyle: tips = { "SleepQuality": "Consider improving sleep hygiene: consistent bedtime, reduce screens before bed, and limit caffeine after noon.", "EducationLevel": "Engaging in lifelong learning (e.g., courses, puzzles, reading) may help build cognitive reserve.", "CardiovascularDisease": "Managing heart health through diet, exercise, and medication (as advised by your doctor) supports brain health.", "AlcoholConsumption": "Moderate or abstinent alcohol use is linked to lower dementia risk; discuss safe limits with your clinician.", "Diabetes": "Keeping blood sugar well-controlled reduces vascular damage to the brain.", "PhysicalActivity": "Aim for 150 mins/week of moderate activity (e.g., brisk walking) to support brain blood flow.", "DietQuality": "A Mediterranean-style diet (vegetables, fish, olive oil, nuts) is associated with lower risk.", "Depression": "Treating depression (therapy/medication) may reduce long-term cognitive decline.", "HeadInjury": "Prevent falls and use helmets during high-risk activities.", "Hypertension": "Regular BP monitoring and management (salt reduction, medication) protect small brain vessels.", "Smoking": "Quitting smoking at any age improves circulation and reduces inflammation.", "BMI": "Maintaining a healthy weight supports metabolic and vascular brain health.", "FamilyHistoryAlzheimers": "While not modifiable, early lifestyle changes can delay onset even with genetic risk.", } tip = tips.get(feat_name_raw, "Discuss personalized strategies with your healthcare provider.") explanation = f"{AUDIO_FEATURE_DESCRIPTIONS.get(feat_name_raw, feat_name_raw) if not is_lifestyle else feat_name_raw} (value = {feat_val}) increased risk ({impact_val}). {tip}" else: desc = AUDIO_FEATURE_DESCRIPTIONS.get(feat_name_raw, "an acoustic speech feature") explanation = f"{desc} (value = {feat_val}) contributed {impact_val} to the risk prediction." explanations.append(explanation[:350]) except Exception: features.append("Unknown") explanations.append(line) return { "topfeatures": features[:5], "explanations": explanations[:5] } # Extract the json from the returned string that the llm gives, as the llm is stupid. def parse_json_from_markdown(s: str) -> dict: """ Extracts and parses JSON from a string that may be wrapped in ``` or ```json. Returns dict on success, or raises ValueError on failure. """ # Try to find JSON inside ```json ... ``` or ``` ... ``` match = re.search(r"```(?:json)?\s*```", s, re.DOTALL | re.IGNORECASE) if match: json_str = match.group(1) else: # Fallback: assume whole string is JSON (e.g., raw dict from function_calling=True) json_str = s.strip() # Parse return json.loads(json_str) # ================================================================ # Crews # ================================================================ # Lifestyle predicton crew def run_crewai_lifestyle(docID: str) -> dict: try: lifestyle_data = pull_lifestyle(docID) if not lifestyle_data: return {"status": "error", "message": f"No lifestyle data for {docID}"} risk_result = calculate_dementia_risk_lifestyle.run(docID) if risk_result["status"] != "success": return risk_result shap_raw = explain_dementia_risk_lifestyle.run(docID) shap_structured = parse_shap_explanation(shap_raw, is_lifestyle=True) # Optional: try crew summary (fallback only) crew_summary = "Summary unavailable." try: risk_assessment_task_lifestyle.description = f"Calculate dementia risk for patient using lifestyle data for ID: {docID}" explanation_task_lifestyle.description = f"Explain the dementia risk prediction for patient {docID} using SHAP" crew = Crew( agents=[risk_calculator_lifestyle], tasks=[risk_assessment_task_lifestyle, explanation_task_lifestyle], verbose=False, function_calling=False ) out = crew.kickoff() if out and str(out).strip(): crew_summary = str(out).strip() except: pass # Build final output — with nested `shap` return { "status": "success", "risk_level": risk_result["risk_level"], "prediction": risk_result["prediction"], "confidence": risk_result["confidence"], "age": risk_result["age"], "gender": risk_result["gender"], "docID": risk_result["docID"], "shap": shap_structured } except Exception as e: return {"status": "error", "message": str(e)} # Audio prediction crew def run_crewai_audio(docID: str) -> dict: try: audio_data = pull_audio(docID) if audio_data is None: return {"status": "error", "message": f"No audio data for {docID}"} risk_result = calculate_dementia_risk_audio.run(docID) if risk_result["status"] != "success": return risk_result shap_raw = explain_dementia_risk_audio.run(docID) shap_structured = parse_shap_explanation(shap_raw, is_lifestyle=False) # Optional crew summary (fallback) crew_summary = "Summary unavailable." try: risk_assessment_task_audio.description = f"Calculate dementia risk for patient using audio data for ID: {docID}" explanation_task_audio.description = f"Explain the dementia risk prediction for patient {docID} using SHAP" crew = Crew( agents=[risk_calculator_audio], tasks=[risk_assessment_task_audio, explanation_task_audio], verbose=False, function_calling=False ) out = crew.kickoff() if out and str(out).strip(): crew_summary = str(out).strip() except: pass return { "status": "success", "risk_level": risk_result["risk_level"], "prediction": risk_result["prediction"], "confidence": risk_result["confidence"], "age": risk_result["age"], "gender": risk_result["gender"], "docID": risk_result["docID"], "shap": shap_structured } except Exception as e: return {"status": "error", "message": str(e)} # Patient news crew def run_crewai_news_patient(count: int = 5) -> dict: crew = Crew( agents=[patient_news_agent], tasks=[get_news_task_patient], verbose=True, function_calling=True ) result = crew.kickoff(inputs={"count": count}) # CrewAI returns ToolOutput → .raw or str — parse if needed try: return parse_json_from_markdown(str(result)) except Exception as e: return {"status": "error", "message": f"Failed to parse crew output, <{e}>", "raw": result} # Doctor news crew def run_crewai_news_doctor(count: int = 5) -> dict: crew = Crew( agents=[doctor_news_agent], tasks=[get_news_task_doctor], verbose=True, function_calling=False ) result = crew.kickoff(inputs={"count": count}) # CrewAI returns ToolOutput → .raw or str — parse if needed try: return parse_json_from_markdown(str(result)) except Exception as e: return {"status": "error", "message": f"Failed to parse crew output, <{e}>", "raw": result}