Compare commits
11 Commits
| Author | SHA1 | Date | |
|---|---|---|---|
| 816c50e467 | |||
| 118e3e63b3 | |||
| 99862629b8 | |||
| 9cc80cd3e6 | |||
| 424d38ad1c | |||
| f1d22b28ad | |||
| 8e4a43c557 | |||
| 2f507bcf20 | |||
| f4bf37f71c | |||
| bc63d1ee72 | |||
| c2ccb8cd11 |
4
.gitignore
vendored
4
.gitignore
vendored
@@ -6,7 +6,11 @@
|
|||||||
.env
|
.env
|
||||||
__pycache__/
|
__pycache__/
|
||||||
*.pyc
|
*.pyc
|
||||||
|
*.csv
|
||||||
|
=======
|
||||||
/reference/
|
/reference/
|
||||||
|
*.svg
|
||||||
|
>>>>>>> Stashed changes
|
||||||
# 2. Ignore virtual environments COMPLETELY
|
# 2. Ignore virtual environments COMPLETELY
|
||||||
# This must come BEFORE the unignore rule
|
# This must come BEFORE the unignore rule
|
||||||
env*/
|
env*/
|
||||||
|
|||||||
600
certainty.py
Normal file
600
certainty.py
Normal file
@@ -0,0 +1,600 @@
|
|||||||
|
|
||||||
|
# %% API call1
|
||||||
|
#import time
|
||||||
|
#import json
|
||||||
|
#import os
|
||||||
|
#from datetime import datetime
|
||||||
|
#import pandas as pd
|
||||||
|
#from openai import OpenAI
|
||||||
|
#from dotenv import load_dotenv
|
||||||
|
#
|
||||||
|
## Load environment variables
|
||||||
|
#load_dotenv()
|
||||||
|
#
|
||||||
|
## === CONFIGURATION ===
|
||||||
|
#OPENAI_API_KEY = os.getenv("OPENAI_API_KEY")
|
||||||
|
#OPENAI_BASE_URL = os.getenv("OPENAI_BASE_URL")
|
||||||
|
#MODEL_NAME = "GPT-OSS-120B"
|
||||||
|
#HEALTH_URL = f"{OPENAI_BASE_URL}/health" # Placeholder - actual health check would need to be implemented
|
||||||
|
#CHAT_URL = f"{OPENAI_BASE_URL}/chat/completions"
|
||||||
|
#
|
||||||
|
## File paths
|
||||||
|
#INPUT_CSV = "/home/shahin/Lab/Doktorarbeit/Barcelona/Data/MS_Briefe_400_with_unique_id_SHA3_explore_cleaned_unique.csv"
|
||||||
|
#EDSS_INSTRUCTIONS_PATH = "/home/shahin/Lab/Doktorarbeit/Barcelona/attach/Komplett.txt"
|
||||||
|
##GRAMMAR_FILE = "/home/shahin/Lab/Doktorarbeit/Barcelona/attach/just_edss_schema.gbnf"
|
||||||
|
#
|
||||||
|
## Initialize OpenAI client
|
||||||
|
#client = OpenAI(
|
||||||
|
# api_key=OPENAI_API_KEY,
|
||||||
|
# base_url=OPENAI_BASE_URL
|
||||||
|
#)
|
||||||
|
#
|
||||||
|
## Read EDSS instructions from file
|
||||||
|
#with open(EDSS_INSTRUCTIONS_PATH, 'r') as f:
|
||||||
|
# EDSS_INSTRUCTIONS = f.read().strip()
|
||||||
|
## === RUN INFERENCE 2 ===
|
||||||
|
#def run_inference(patient_text):
|
||||||
|
# prompt = f'''
|
||||||
|
# Du bist ein medizinischer Assistent, der spezialisiert darauf ist, EDSS-Scores (Expanded Disability Status Scale) aus klinischen Berichten zu extrahieren.
|
||||||
|
#### Regeln für die Ausgabe:
|
||||||
|
#1. **Reason**: Erstelle eine prägnante Zusammenfassung (max. 400 Zeichen) der Befunde auf **DEUTSCH**, die zur Einstufung führen.
|
||||||
|
#2. **klassifizierbar**:
|
||||||
|
# - Setze dies auf **true**, wenn ein EDSS-Wert identifiziert, berechnet oder basierend auf den klinischen Hinweisen plausibel geschätzt werden kann.
|
||||||
|
# - Setze dies auf **false**, NUR wenn die Daten absolut unzureichend oder so widersprüchlich sind, dass keinerlei Einstufung möglich ist.
|
||||||
|
#3. **EDSS**:
|
||||||
|
# - Dieses Feld ist **VERPFLICHTEND**, wenn "klassifizierbar" auf true steht.
|
||||||
|
# - Es muss eine Zahl zwischen 0.0 und 10.0 sein.
|
||||||
|
# - Versuche stets, den EDSS-Wert so präzise wie möglich zu bestimmen, auch wenn die Datenlage dünn ist (nutze verfügbare Informationen zu Gehstrecke und Funktionssystemen).
|
||||||
|
# - Dieses Feld **DARF NICHT ERSCHEINEN**, wenn "klassifizierbar" auf false steht.
|
||||||
|
#
|
||||||
|
#### Einschränkungen:
|
||||||
|
#- Erfinde keine Fakten, aber nutze klinische Herleitungen aus dem Bericht, um den EDSS zu bestimmen.
|
||||||
|
#- Priorisiere die Vergabe eines EDSS-Wertes gegenüber der Markierung als nicht klassifizierbar.
|
||||||
|
#- Halte dich strikt an die JSON-Struktur.
|
||||||
|
#
|
||||||
|
#EDSS-Bewertungsrichtlinien:
|
||||||
|
#{EDSS_INSTRUCTIONS}
|
||||||
|
#
|
||||||
|
#Patientenbericht:
|
||||||
|
#{patient_text}
|
||||||
|
#'''
|
||||||
|
# start_time = time.time()
|
||||||
|
#
|
||||||
|
# try:
|
||||||
|
# # Make API call using OpenAI client
|
||||||
|
# response = client.chat.completions.create(
|
||||||
|
# messages=[
|
||||||
|
# {
|
||||||
|
# "role": "system",
|
||||||
|
# "content": "You extract EDSS scores. You prioritize providing a score even if data is partial, by using clinical inference."
|
||||||
|
# },
|
||||||
|
# {
|
||||||
|
# "role": "user",
|
||||||
|
# "content": prompt
|
||||||
|
# }
|
||||||
|
# ],
|
||||||
|
# model=MODEL_NAME,
|
||||||
|
# max_tokens=2048,
|
||||||
|
# temperature=0.0,
|
||||||
|
# response_format={"type": "json_object"}
|
||||||
|
# )
|
||||||
|
#
|
||||||
|
# # Extract content from response
|
||||||
|
# content = response.choices[0].message.content
|
||||||
|
#
|
||||||
|
# # Parse the JSON response
|
||||||
|
# parsed = json.loads(content)
|
||||||
|
#
|
||||||
|
# inference_time = time.time() - start_time
|
||||||
|
#
|
||||||
|
# return {
|
||||||
|
# "success": True,
|
||||||
|
# "result": parsed,
|
||||||
|
# "inference_time_sec": inference_time
|
||||||
|
# }
|
||||||
|
#
|
||||||
|
# except Exception as e:
|
||||||
|
# print(f"Inference error: {e}")
|
||||||
|
# return {
|
||||||
|
# "success": False,
|
||||||
|
# "error": str(e),
|
||||||
|
# "inference_time_sec": -1
|
||||||
|
# }
|
||||||
|
## === BUILD PATIENT TEXT ===
|
||||||
|
#def build_patient_text(row):
|
||||||
|
# return (
|
||||||
|
# str(row["T_Zusammenfassung"]) + "\n" +
|
||||||
|
# str(row["Diagnosen"]) + "\n" +
|
||||||
|
# str(row["T_KlinBef"]) + "\n" +
|
||||||
|
# str(row["T_Befunde"]) + "\n"
|
||||||
|
# )
|
||||||
|
#
|
||||||
|
#if __name__ == "__main__":
|
||||||
|
# # Read CSV file ONLY inside main block
|
||||||
|
# df = pd.read_csv(INPUT_CSV, sep=';')
|
||||||
|
# results = []
|
||||||
|
#
|
||||||
|
# # Process each row
|
||||||
|
# for idx, row in df.iterrows():
|
||||||
|
# print(f"Processing row {idx + 1}/{len(df)}")
|
||||||
|
# try:
|
||||||
|
# patient_text = build_patient_text(row)
|
||||||
|
# result = run_inference(patient_text)
|
||||||
|
#
|
||||||
|
# # Add unique_id and MedDatum to result for tracking
|
||||||
|
# result["unique_id"] = row.get("unique_id", f"row_{idx}")
|
||||||
|
# result["MedDatum"] = row.get("MedDatum", None)
|
||||||
|
#
|
||||||
|
# results.append(result)
|
||||||
|
# print(json.dumps(result, indent=2))
|
||||||
|
# except Exception as e:
|
||||||
|
# print(f"Error processing row {idx}: {e}")
|
||||||
|
# results.append({
|
||||||
|
# "success": False,
|
||||||
|
# "error": str(e),
|
||||||
|
# "unique_id": row.get("unique_id", f"row_{idx}"),
|
||||||
|
# "MedDatum": row.get("MedDatum", None)
|
||||||
|
# })
|
||||||
|
#
|
||||||
|
# # Save results to a JSON file
|
||||||
|
# output_json = INPUT_CSV.replace(".csv", "_results_Nisch.json")
|
||||||
|
# with open(output_json, 'w') as f:
|
||||||
|
# json.dump(results, f, indent=2)
|
||||||
|
# print(f"Results saved to {output_json}")
|
||||||
|
##
|
||||||
|
|
||||||
|
|
||||||
|
|
||||||
|
# %% API call1 - Enhanced with certainty scoring
|
||||||
|
#import time
|
||||||
|
#import json
|
||||||
|
#import os
|
||||||
|
#from datetime import datetime
|
||||||
|
#import pandas as pd
|
||||||
|
#from openai import OpenAI
|
||||||
|
#from dotenv import load_dotenv
|
||||||
|
#
|
||||||
|
## Load environment variables
|
||||||
|
#load_dotenv()
|
||||||
|
#
|
||||||
|
## === CONFIGURATION ===
|
||||||
|
#OPENAI_API_KEY = os.getenv("OPENAI_API_KEY")
|
||||||
|
#OPENAI_BASE_URL = os.getenv("OPENAI_BASE_URL")
|
||||||
|
#MODEL_NAME = "GPT-OSS-120B"
|
||||||
|
#
|
||||||
|
## File paths
|
||||||
|
#INPUT_CSV = "/home/shahin/Lab/Doktorarbeit/Barcelona/Data/Test.csv"
|
||||||
|
#EDSS_INSTRUCTIONS_PATH = "/home/shahin/Lab/Doktorarbeit/Barcelona/attach/Komplett.txt"
|
||||||
|
#
|
||||||
|
## Initialize OpenAI client
|
||||||
|
#client = OpenAI(
|
||||||
|
# api_key=OPENAI_API_KEY,
|
||||||
|
# base_url=OPENAI_BASE_URL
|
||||||
|
#)
|
||||||
|
#
|
||||||
|
## Read EDSS instructions from file
|
||||||
|
#with open(EDSS_INSTRUCTIONS_PATH, 'r') as f:
|
||||||
|
# EDSS_INSTRUCTIONS = f.read().strip()
|
||||||
|
#
|
||||||
|
## === PROMPT WITH CERTAINTY REQUEST ===
|
||||||
|
#def build_prompt(patient_text):
|
||||||
|
# return f'''Du bist ein medizinischer Assistent, der spezialisiert darauf ist, EDSS-Scores (Expanded Disability Status Scale), alle Unterkategorien und die Bewertungssicherheit aus klinischen Berichten zu extrahieren.
|
||||||
|
#
|
||||||
|
#### Deine Aufgabe:
|
||||||
|
#1. Analysiere den Patientenbericht und extrahiere:
|
||||||
|
# - Den Gesamt-EDSS-Score (0.0–10.0)
|
||||||
|
# - Alle 8 EDSS-Unterkategorien (mit jeweils eigener Maximalpunktzahl)
|
||||||
|
#2. Schätze für jede Entscheidung die Sicherheit als Ganzzahl von 0–100 % ein.
|
||||||
|
#
|
||||||
|
#### Struktur der JSON-Ausgabe (VERPFLICHTEND):
|
||||||
|
#Gib NUR gültiges JSON zurück — kein Markdown, kein Text davor/dahinter.
|
||||||
|
#
|
||||||
|
#{{
|
||||||
|
# "reason": "Kernaussage zur EDSS-Begründung (max. 400 Zeichen, auf Deutsch).",
|
||||||
|
# "klassifizierbar": true/false,
|
||||||
|
# "EDSS": null ODER Zahl zwischen 0.0 und 10.0 (nur wenn klassifizierbar=true)",
|
||||||
|
# "certainty_percent": 0 ODER Zahl zwischen 0 und 100 (Ganzzahl)",
|
||||||
|
# "subcategories": {{
|
||||||
|
# "VISUAL_OPTIC_FUNCTIONS": null ODER Zahl zwischen 0.0 und 6.0,
|
||||||
|
# "BRAINSTEM_FUNCTIONS": null ODER Zahl zwischen 0.0 und 6.0,
|
||||||
|
# "PYRAMIDAL_FUNCTIONS": null ODER Zahl zwischen 0.0 und 6.0,
|
||||||
|
# "CEREBELLAR_FUNCTIONS": null ODER Zahl zwischen 0.0 und 6.0,
|
||||||
|
# "SENSORY_FUNCTIONS": null ODER Zahl zwischen 0.0 und 6.0,
|
||||||
|
# "BOWEL_AND_BLADDER_FUNCTIONS": null ODER Zahl zwischen 0.0 und 6.0,
|
||||||
|
# "CEREBRAL_FUNCTIONS": null ODER Zahl zwischen 0.0 und 6.0,
|
||||||
|
# "AMBULATION": null ODER Zahl zwischen 0.0 und 10.0
|
||||||
|
# }}
|
||||||
|
#}}
|
||||||
|
#
|
||||||
|
#### Regeln:
|
||||||
|
#- **reason**: Kurze, prägnante Begründung (auf Deutsch, max. 400 Zeichen), warum du den EDSS-Wert und die Unterkategorien so bewertest.
|
||||||
|
#- **klassifizierbar**:
|
||||||
|
# - `true`, wenn EDSS und mindestens die wichtigsten Unterkategorien *eindeutig ableitbar* oder *plausibel inferierbar* sind.
|
||||||
|
# - `false`, **nur**, wenn keine relevanten Daten vorliegen, oder diese so widersprüchlich/inkonsistent sind, dass keine vernünftige Einschätzung möglich ist.
|
||||||
|
#- **EDSS**:
|
||||||
|
# - **VERPFLICHTEND**, wenn `klassifizierbar=true`.
|
||||||
|
# - Zahl zwischen 0.0 und 10.0 (z.B. 3.0, 5.5). Darf **nicht** erscheinen, wenn `klassifizierbar=false`.
|
||||||
|
#- **certainty_percent**:
|
||||||
|
# - **Immer present** — Ganzzahl (0–100), basierend auf:
|
||||||
|
# - Klarheit und Vollständigkeit der Berichtsangaben,
|
||||||
|
# - Stichhaltigkeit der Schlussfolgerung (inkl. Inferenz),
|
||||||
|
# - Konsistenz zwischen den Unterkategorien.
|
||||||
|
#- **subcategories**:
|
||||||
|
# - **Immer present** — **alle 8 Unterkategorien** müssen enthalten sein.
|
||||||
|
# - Jeder Wert ist entweder:
|
||||||
|
# - `null` (wenn keine ausreichende Information vorliegt), **oder**
|
||||||
|
# - eine Zahl ≤ jeweiliger Obergrenze (z.B. Ambulation ≤ 10.0).
|
||||||
|
# - Wenn die Unterkategorie plausibel inferiert werden kann (auch indirekt), gib einen sinnvollen Wert ab.
|
||||||
|
# - Beispiel: Wenn „Gang mit Krückstock auf ebenem Boden bis 200 m“ steht, setze `AMBULATION: 5.5`.
|
||||||
|
#
|
||||||
|
#### EDSS-Bewertungsrichtlinien:
|
||||||
|
#{EDSS_INSTRUCTIONS}
|
||||||
|
#
|
||||||
|
#Patientenbericht:
|
||||||
|
#{patient_text}
|
||||||
|
#'''
|
||||||
|
#
|
||||||
|
## === INFERENCE FUNCTION ===
|
||||||
|
#def run_inference(patient_text):
|
||||||
|
# prompt = build_prompt(patient_text)
|
||||||
|
#
|
||||||
|
# start_time = time.time()
|
||||||
|
#
|
||||||
|
# try:
|
||||||
|
# response = client.chat.completions.create(
|
||||||
|
# messages=[
|
||||||
|
# {"role": "system", "content": "Du gibst EXKLUSIV gültiges JSON zurück — keine weiteren Erklärungen."}
|
||||||
|
# ] + [
|
||||||
|
# {"role": "user", "content": prompt}
|
||||||
|
# ],
|
||||||
|
# model=MODEL_NAME,
|
||||||
|
# max_tokens=2048,
|
||||||
|
# temperature=0.1, # Slightly higher for more natural certainty estimation (still low for reliability)
|
||||||
|
# response_format={"type": "json_object"}
|
||||||
|
# )
|
||||||
|
#
|
||||||
|
# content = response.choices[0].message.content
|
||||||
|
#
|
||||||
|
# # Parse and validate JSON
|
||||||
|
# try:
|
||||||
|
# parsed = json.loads(content)
|
||||||
|
# except json.JSONDecodeError as e:
|
||||||
|
# print(f"⚠️ JSON parsing failed: {e}")
|
||||||
|
# print("Raw response:", content[:500])
|
||||||
|
# raise ValueError("Model did not return valid JSON")
|
||||||
|
#
|
||||||
|
# # Enforce required keys
|
||||||
|
# if "certainty_percent" not in parsed:
|
||||||
|
# print("⚠️ Missing 'certainty_percent' in output! Force-adding fallback.")
|
||||||
|
# parsed["certainty_percent"] = 0 # fallback
|
||||||
|
# elif not isinstance(parsed["certainty_percent"], (int, float)):
|
||||||
|
# parsed["certainty_percent"] = int(parsed["certainty_percent"])
|
||||||
|
#
|
||||||
|
# # Clamp certainty to [0, 100]
|
||||||
|
# pct = parsed["certainty_percent"]
|
||||||
|
# parsed["certainty_percent"] =max(0, min(100, int(pct)))
|
||||||
|
#
|
||||||
|
# # Enforce EDSS rules: if not classifiable → remove EDSS
|
||||||
|
# if not parsed.get("klassifizierbar", False):
|
||||||
|
# if "EDSS" in parsed:
|
||||||
|
# del parsed["EDSS"] # per spec, must not appear if not classifiable
|
||||||
|
# else:
|
||||||
|
# if "EDSS" not in parsed:
|
||||||
|
# print("⚠️ 'klassifizierbar' is true but EDSS missing — adding fallback.")
|
||||||
|
# parsed["EDSS"] = 7.0 # last-resort fallback
|
||||||
|
#
|
||||||
|
# inference_time = time.time() - start_time
|
||||||
|
#
|
||||||
|
# return {
|
||||||
|
# "success": True,
|
||||||
|
# "result": parsed,
|
||||||
|
# "inference_time_sec": inference_time
|
||||||
|
# }
|
||||||
|
#
|
||||||
|
# except Exception as e:
|
||||||
|
# print(f"❌ Inference error: {e}")
|
||||||
|
# return {
|
||||||
|
# "success": False,
|
||||||
|
# "error": str(e),
|
||||||
|
# "inference_time_sec": -1,
|
||||||
|
# "result": None # no structured output
|
||||||
|
# }
|
||||||
|
#
|
||||||
|
## === BUILD PATIENT TEXT ===
|
||||||
|
#def build_patient_text(row):
|
||||||
|
# return (
|
||||||
|
# str(row.get("T_Zusammenfassung", "")) + "\n" +
|
||||||
|
# str(row.get("Diagnosen", "")) + "\n" +
|
||||||
|
# str(row.get("T_KlinBef", "")) + "\n" +
|
||||||
|
# str(row.get("T_Befunde", ""))
|
||||||
|
# )
|
||||||
|
#
|
||||||
|
#if __name__ == "__main__":
|
||||||
|
# # Load data
|
||||||
|
# df = pd.read_csv(INPUT_CSV, sep=';')
|
||||||
|
# results = []
|
||||||
|
#
|
||||||
|
# # Optional: limit for testing
|
||||||
|
# # df = df.head(3)
|
||||||
|
#
|
||||||
|
# print(f"Processing {len(df)} rows...")
|
||||||
|
# for idx, row in df.iterrows():
|
||||||
|
# print(f"\n— Row {idx + 1}/{len(df)} —")
|
||||||
|
# try:
|
||||||
|
# patient_text = build_patient_text(row)
|
||||||
|
# result = run_inference(patient_text)
|
||||||
|
#
|
||||||
|
# # Attach metadata
|
||||||
|
# result["unique_id"] = row.get("unique_id", f"row_{idx}")
|
||||||
|
# result["MedDatum"] = row.get("MedDatum", None)
|
||||||
|
#
|
||||||
|
# results.append(result)
|
||||||
|
#
|
||||||
|
# # Print summary
|
||||||
|
# if result["success"]:
|
||||||
|
# res = result["result"]
|
||||||
|
# edss = res.get("EDSS", "N/A") if res.get("klassifizierbar") else "N/A"
|
||||||
|
# print(f"✅ Result → EDSS={edss}, certainty={res.get('certainty_percent', 'N/A')}%")
|
||||||
|
# print(f" Reason: {res.get('reason', 'N/A')[:100]}…")
|
||||||
|
# else:
|
||||||
|
# print(f"❌ Failed: {result.get('error', 'Unknown error')[:100]}")
|
||||||
|
#
|
||||||
|
# except Exception as e:
|
||||||
|
# print(f"⚠️ Error processing row {idx}: {e}")
|
||||||
|
# results.append({
|
||||||
|
# "success": False,
|
||||||
|
# "error": str(e),
|
||||||
|
# "unique_id": row.get("unique_id", f"row_{idx}"),
|
||||||
|
# "MedDatum": row.get("MedDatum", None),
|
||||||
|
# "result": None
|
||||||
|
# })
|
||||||
|
#
|
||||||
|
# # Save results
|
||||||
|
# output_json = INPUT_CSV.replace(".csv", "_results_Nisch_certainty.json")
|
||||||
|
# with open(output_json, 'w', encoding='utf-8') as f:
|
||||||
|
# json.dump(results, f, indent=2, ensure_ascii=False)
|
||||||
|
# print(f"\n✅ Saved results to: {output_json}")
|
||||||
|
#
|
||||||
|
##
|
||||||
|
|
||||||
|
|
||||||
|
# %% API call - Multi-iteration EDSS + certainty extraction
|
||||||
|
|
||||||
|
import time
|
||||||
|
import json
|
||||||
|
import os
|
||||||
|
from datetime import datetime
|
||||||
|
import pandas as pd
|
||||||
|
from openai import OpenAI
|
||||||
|
from dotenv import load_dotenv
|
||||||
|
|
||||||
|
# Load environment variables
|
||||||
|
load_dotenv()
|
||||||
|
|
||||||
|
# === CONFIGURATION ===
|
||||||
|
OPENAI_API_KEY = os.getenv("OPENAI_API_KEY")
|
||||||
|
OPENAI_BASE_URL = os.getenv("OPENAI_BASE_URL")
|
||||||
|
MODEL_NAME = "GPT-OSS-120B"
|
||||||
|
|
||||||
|
# File paths
|
||||||
|
INPUT_CSV = "/home/shahin/Lab/Doktorarbeit/Barcelona/Data/MS_Briefe_400_with_unique_id_SHA3_explore_cleaned_unique.csv"
|
||||||
|
EDSS_INSTRUCTIONS_PATH = "/home/shahin/Lab/Doktorarbeit/Barcelona/attach/Komplett.txt"
|
||||||
|
|
||||||
|
# Iteration settings
|
||||||
|
NUM_ITERATIONS = 20
|
||||||
|
STOP_ON_FIRST_ERROR = False # Set to True for debugging
|
||||||
|
|
||||||
|
# Initialize OpenAI client
|
||||||
|
client = OpenAI(
|
||||||
|
api_key=OPENAI_API_KEY,
|
||||||
|
base_url=OPENAI_BASE_URL
|
||||||
|
)
|
||||||
|
|
||||||
|
# Read EDSS instructions from file
|
||||||
|
with open(EDSS_INSTRUCTIONS_PATH, 'r') as f:
|
||||||
|
EDSS_INSTRUCTIONS = f.read().strip()
|
||||||
|
|
||||||
|
# === PROMPT (unchanged from before) ===
|
||||||
|
def build_prompt(patient_text):
|
||||||
|
return f'''Du bist ein medizinischer Assistent, der spezialisiert darauf ist, EDSS-Scores (Expanded Disability Status Scale), alle Unterkategorien und die Bewertungssicherheit aus klinischen Berichten zu extrahieren.
|
||||||
|
|
||||||
|
### Deine Aufgabe:
|
||||||
|
1. Analysiere den Patientenbericht und extrahiere:
|
||||||
|
- Den Gesamt-EDSS-Score (0.0–10.0)
|
||||||
|
- Alle 8 EDSS-Unterkategorien (mit jeweils eigener Maximalpunktzahl)
|
||||||
|
2. Schätze für jede Entscheidung die Sicherheit als Ganzzahl von 0–100 % ein.
|
||||||
|
|
||||||
|
### Struktur der JSON-Ausgabe (VERPFLICHTEND):
|
||||||
|
Gib NUR gültiges JSON zurück — kein Markdown, kein Text davor/dahinter.
|
||||||
|
|
||||||
|
{{
|
||||||
|
"reason": "Kernaussage zur EDSS-Begründung (max. 400 Zeichen, auf Deutsch).",
|
||||||
|
"klassifizierbar": true/false,
|
||||||
|
"EDSS": null ODER Zahl zwischen 0.0 und 10.0 (nur wenn klassifizierbar=true)",
|
||||||
|
"certainty_percent": 0 ODER Zahl zwischen 0 und 100 (Ganzzahl)",
|
||||||
|
"subcategories": {{
|
||||||
|
"VISUAL_OPTIC_FUNCTIONS": null ODER Zahl zwischen 0.0 und 6.0,
|
||||||
|
"BRAINSTEM_FUNCTIONS": null ODER Zahl zwischen 0.0 und 6.0,
|
||||||
|
"PYRAMIDAL_FUNCTIONS": null ODER Zahl zwischen 0.0 und 6.0,
|
||||||
|
"CEREBELLAR_FUNCTIONS": null ODER Zahl zwischen 0.0 und 6.0,
|
||||||
|
"SENSORY_FUNCTIONS": null ODER Zahl zwischen 0.0 und 6.0,
|
||||||
|
"BOWEL_AND_BLADDER_FUNCTIONS": null ODER Zahl zwischen 0.0 und 6.0,
|
||||||
|
"CEREBRAL_FUNCTIONS": null ODER Zahl zwischen 0.0 und 6.0,
|
||||||
|
"AMBULATION": null ODER Zahl zwischen 0.0 und 10.0
|
||||||
|
}}
|
||||||
|
}}
|
||||||
|
|
||||||
|
### Regeln:
|
||||||
|
- **reason**: Kurze, prägnante Begründung (auf Deutsch, max. 400 Zeichen), warum du den EDSS-Wert und die Unterkategorien so bewertest.
|
||||||
|
- **klassifizierbar**:
|
||||||
|
- `true`, wenn EDSS und mindestens die wichtigsten Unterkategorien *eindeutig ableitbar* oder *plausibel inferierbar* sind.
|
||||||
|
- `false`, **nur**, wenn keine relevanten Daten vorliegen, oder diese so widersprüchlich/inkonsistent sind, dass keine vernünftige Einschätzung möglich ist.
|
||||||
|
- **EDSS**:
|
||||||
|
- **VERPFLICHTEND**, wenn `klassifizierbar=true`.
|
||||||
|
- Zahl zwischen 0.0 und 10.0 (z.B. 3.0, 5.5). Darf **nicht** erscheinen, wenn `klassifizierbar=false`.
|
||||||
|
- **certainty_percent**:
|
||||||
|
- **Immer present** — Ganzzahl (0–100), basierend auf:
|
||||||
|
- Klarheit und Vollständigkeit der Berichtsangaben,
|
||||||
|
- Stichhaltigkeit der Schlussfolgerung (inkl. Inferenz),
|
||||||
|
- Konsistenz zwischen den Unterkategorien.
|
||||||
|
- **subcategories**:
|
||||||
|
- **Immer present** — **alle 8 Unterkategorien** müssen enthalten sein.
|
||||||
|
- Jeder Wert ist entweder:
|
||||||
|
- `null` (wenn keine ausreichende Information vorliegt), **oder**
|
||||||
|
- eine Zahl ≤ jeweiliger Obergrenze (z.B. Ambulation ≤ 10.0).
|
||||||
|
- Wenn die Unterkategorie plausibel inferiert werden kann (auch indirekt), gib einen sinnvollen Wert ab.
|
||||||
|
- Beispiel: Wenn „Gang mit Krückstock auf ebenem Boden bis 200 m“ steht, setze `AMBULATION: 5.5`.
|
||||||
|
|
||||||
|
### EDSS-Bewertungsrichtlinien:
|
||||||
|
{EDSS_INSTRUCTIONS}
|
||||||
|
|
||||||
|
Patientenbericht:
|
||||||
|
{patient_text}
|
||||||
|
'''
|
||||||
|
|
||||||
|
# === INFERENCE FUNCTION (unchanged) ===
|
||||||
|
def run_inference(patient_text):
|
||||||
|
prompt = build_prompt(patient_text)
|
||||||
|
|
||||||
|
start_time = time.time()
|
||||||
|
|
||||||
|
try:
|
||||||
|
response = client.chat.completions.create(
|
||||||
|
messages=[
|
||||||
|
{"role": "system", "content": "Du gibst EXKLUSIV gültiges JSON zurück — keine weiteren Erklärungen."}
|
||||||
|
] + [
|
||||||
|
{"role": "user", "content": prompt}
|
||||||
|
],
|
||||||
|
model=MODEL_NAME,
|
||||||
|
max_tokens=2048,
|
||||||
|
temperature=0.1,
|
||||||
|
response_format={"type": "json_object"}
|
||||||
|
)
|
||||||
|
|
||||||
|
content = response.choices[0].message.content
|
||||||
|
|
||||||
|
# Parse and validate JSON
|
||||||
|
try:
|
||||||
|
parsed = json.loads(content)
|
||||||
|
except json.JSONDecodeError as e:
|
||||||
|
print(f"⚠️ JSON parsing failed: {e}")
|
||||||
|
print("Raw response:", content[:500])
|
||||||
|
raise ValueError("Model did not return valid JSON")
|
||||||
|
|
||||||
|
# Enforce required keys
|
||||||
|
if "certainty_percent" not in parsed:
|
||||||
|
print("⚠️ Missing 'certainty_percent' in output! Force-adding fallback.")
|
||||||
|
parsed["certainty_percent"] = 0
|
||||||
|
elif not isinstance(parsed["certainty_percent"], (int, float)):
|
||||||
|
parsed["certainty_percent"] = int(parsed["certainty_percent"])
|
||||||
|
|
||||||
|
# Clamp certainty to [0, 100]
|
||||||
|
pct = parsed["certainty_percent"]
|
||||||
|
parsed["certainty_percent"] = max(0, min(100, int(pct)))
|
||||||
|
|
||||||
|
# Enforce EDSS rules
|
||||||
|
if not parsed.get("klassifizierbar", False):
|
||||||
|
if "EDSS" in parsed:
|
||||||
|
del parsed["EDSS"]
|
||||||
|
else:
|
||||||
|
if "EDSS" not in parsed:
|
||||||
|
print("⚠️ 'klassifizierbar' is true but EDSS missing — adding fallback.")
|
||||||
|
parsed["EDSS"] = 7.0
|
||||||
|
|
||||||
|
inference_time = time.time() - start_time
|
||||||
|
|
||||||
|
return {
|
||||||
|
"success": True,
|
||||||
|
"result": parsed,
|
||||||
|
"inference_time_sec": inference_time
|
||||||
|
}
|
||||||
|
|
||||||
|
except Exception as e:
|
||||||
|
print(f"❌ Inference error: {e}")
|
||||||
|
return {
|
||||||
|
"success": False,
|
||||||
|
"error": str(e),
|
||||||
|
"inference_time_sec": -1,
|
||||||
|
"result": None
|
||||||
|
}
|
||||||
|
|
||||||
|
# === BUILD PATIENT TEXT ===
|
||||||
|
def build_patient_text(row):
|
||||||
|
return (
|
||||||
|
str(row.get("T_Zusammenfassung", "")) + "\n" +
|
||||||
|
str(row.get("Diagnosen", "")) + "\n" +
|
||||||
|
str(row.get("T_KlinBef", "")) + "\n" +
|
||||||
|
str(row.get("T_Befunde", ""))
|
||||||
|
)
|
||||||
|
|
||||||
|
# === MAIN LOOP (NEW: MULTI-ITERATION) ===
|
||||||
|
if __name__ == "__main__":
|
||||||
|
# Load data ONCE (to avoid repeated I/O overhead)
|
||||||
|
df = pd.read_csv(INPUT_CSV, sep=';')
|
||||||
|
total_rows = len(df)
|
||||||
|
print(f"Loaded {total_rows} patient records.")
|
||||||
|
|
||||||
|
for iteration in range(1, NUM_ITERATIONS + 1):
|
||||||
|
print(f"\n{'='*60}")
|
||||||
|
print(f"🔄 ITERATION {iteration}/{NUM_ITERATIONS}")
|
||||||
|
print(f"{'='*60}")
|
||||||
|
|
||||||
|
iteration_results = []
|
||||||
|
start_iter = time.time()
|
||||||
|
|
||||||
|
for idx, row in df.iterrows():
|
||||||
|
print(f"\rRow {idx+1}/{total_rows} | Iter {iteration}", end='', flush=True)
|
||||||
|
try:
|
||||||
|
patient_text = build_patient_text(row)
|
||||||
|
result = run_inference(patient_text)
|
||||||
|
|
||||||
|
# Attach metadata
|
||||||
|
if result["success"]:
|
||||||
|
res = result["result"].copy() # avoid mutation
|
||||||
|
res["iteration"] = iteration
|
||||||
|
res["unique_id"] = row.get("unique_id", f"row_{idx}")
|
||||||
|
res["MedDatum"] = row.get("MedDatum", None)
|
||||||
|
result["result"] = res
|
||||||
|
|
||||||
|
else:
|
||||||
|
result["iteration"] = iteration
|
||||||
|
result["unique_id"] = row.get("unique_id", f"row_{idx}")
|
||||||
|
result["MedDatum"] = row.get("MedDatum", None)
|
||||||
|
|
||||||
|
iteration_results.append(result)
|
||||||
|
|
||||||
|
if result["success"]:
|
||||||
|
res = result["result"]
|
||||||
|
edss = res.get("EDSS", "N/A") if res.get("klassifizierbar") else "N/A"
|
||||||
|
print(f" ✅ EDSS={edss}, cert={res.get('certainty_percent', '?')}%")
|
||||||
|
else:
|
||||||
|
print(f" ❌ {result.get('error', 'Unknown')}")
|
||||||
|
|
||||||
|
except Exception as e:
|
||||||
|
print(f"\n⚠️ Row {idx} failed: {e}")
|
||||||
|
iteration_results.append({
|
||||||
|
"success": False,
|
||||||
|
"error": str(e),
|
||||||
|
"iteration": iteration,
|
||||||
|
"unique_id": row.get("unique_id", f"row_{idx}"),
|
||||||
|
"MedDatum": row.get("MedDatum", None),
|
||||||
|
"result": None
|
||||||
|
})
|
||||||
|
if STOP_ON_FIRST_ERROR:
|
||||||
|
break
|
||||||
|
|
||||||
|
# Save per-iteration results
|
||||||
|
timestamp = datetime.now().strftime("%Y%m%d_%H%M%S")
|
||||||
|
output_path = INPUT_CSV.replace(".csv", f"_results_iter_{iteration}_{timestamp}.json")
|
||||||
|
with open(output_path, 'w', encoding='utf-8') as f:
|
||||||
|
json.dump(iteration_results, f, indent=2, ensure_ascii=False)
|
||||||
|
print(f"\n✅ Iteration {iteration} complete. Saved to: {output_path}")
|
||||||
|
|
||||||
|
elapsed = time.time() - start_iter
|
||||||
|
print(f"⏱️ Iteration {iteration} took {elapsed:.1f}s ({elapsed/total_rows:.1f}s/row)")
|
||||||
|
|
||||||
|
print(f"\n🎉 All {NUM_ITERATIONS} iterations completed!")
|
||||||
|
|
||||||
|
|
||||||
|
|
||||||
|
##
|
||||||
1540
certainty_show.py
Normal file
1540
certainty_show.py
Normal file
File diff suppressed because it is too large
Load Diff
117
figure1.py
117
figure1.py
@@ -263,3 +263,120 @@ plt.legend(frameon=False, loc='upper center', bbox_to_anchor=(0.5, -0.05))
|
|||||||
plt.tight_layout()
|
plt.tight_layout()
|
||||||
plt.show()
|
plt.show()
|
||||||
##
|
##
|
||||||
|
|
||||||
|
|
||||||
|
|
||||||
|
|
||||||
|
# %% name
|
||||||
|
import matplotlib.pyplot as plt
|
||||||
|
|
||||||
|
# Data
|
||||||
|
data = {
|
||||||
|
'Visit': [9, 8, 7, 6, 5, 4, 3, 2, 1],
|
||||||
|
'patient_count': [2, 3, 3, 6, 13, 17, 28, 24, 32]
|
||||||
|
}
|
||||||
|
|
||||||
|
# Create figure and axis
|
||||||
|
fig, ax = plt.subplots(figsize=(10, 6))
|
||||||
|
|
||||||
|
# Plot the bar chart
|
||||||
|
bars = ax.bar(data['Visit'], data['patient_count'], color='darkblue', label='Patients by Visit Count')
|
||||||
|
|
||||||
|
# Add labels and title
|
||||||
|
ax.set_xlabel('Visit Number (from last to first)', fontsize=12)
|
||||||
|
ax.set_ylabel('Number of Patients', fontsize=12)
|
||||||
|
ax.set_title('Patient Visits by Visit Number', fontsize=14)
|
||||||
|
|
||||||
|
# Invert x-axis to show Visit 9 on the left (descending order) if desired, but keep natural order (1–9 left to right)
|
||||||
|
# For descending order (9→1 from left to right), we'd need to reverse:
|
||||||
|
# Visit = data['Visit'][::-1], patient_count = data['patient_count'][::-1]
|
||||||
|
# But standard practice is ascending (1 to 9), so we'll sort accordingly:
|
||||||
|
# Let's sort by Visit to ensure left-to-right: 1,2,...,9
|
||||||
|
|
||||||
|
# Actually, your current Visit list is [9,8,...,1], which is descending.
|
||||||
|
# Let's sort by Visit for intuitive left-to-right increasing order:
|
||||||
|
sorted_indices = sorted(range(len(data['Visit'])), key=lambda i: data['Visit'][i])
|
||||||
|
visit_sorted = [data['Visit'][i] for i in sorted_indices]
|
||||||
|
count_sorted = [data['patient_count'][i] for i in sorted_indices]
|
||||||
|
|
||||||
|
# Re-plot with sorted x-axis:
|
||||||
|
ax.clear()
|
||||||
|
bars = ax.bar(visit_sorted, count_sorted, color='darkblue', label='Patients by Visit Count')
|
||||||
|
|
||||||
|
# Re-apply labels, etc.
|
||||||
|
ax.set_xlabel('Number of Visits', fontsize=12)
|
||||||
|
ax.set_ylabel('Number of Unique Patients', fontsize=12)
|
||||||
|
#ax.set_title('Number of Patients by Visit Number', fontsize=14)
|
||||||
|
|
||||||
|
# Add legend
|
||||||
|
ax.legend()
|
||||||
|
|
||||||
|
# Improve layout and grid
|
||||||
|
ax.grid(axis='y', linestyle='--', alpha=0.7)
|
||||||
|
plt.xticks(visit_sorted) # Ensure all integer visit numbers are shown
|
||||||
|
|
||||||
|
# Show the plot
|
||||||
|
plt.tight_layout()
|
||||||
|
plt.show()
|
||||||
|
|
||||||
|
##
|
||||||
|
|
||||||
|
# %% Patientjourney Bubble chart
|
||||||
|
import matplotlib.pyplot as plt
|
||||||
|
import numpy as np
|
||||||
|
|
||||||
|
import matplotlib as mpl
|
||||||
|
|
||||||
|
mpl.rcParams["font.family"] = "DejaVu Sans" # or "Arial", "Calibri", "Times New Roman", ...
|
||||||
|
mpl.rcParams["font.size"] = 12 # default size for text
|
||||||
|
mpl.rcParams["axes.titlesize"] = 14
|
||||||
|
mpl.rcParams["axes.titleweight"] = "bold"
|
||||||
|
|
||||||
|
|
||||||
|
# Data (your counts)
|
||||||
|
visits = np.array([1, 2, 3, 4, 5, 6, 7, 8, 9])
|
||||||
|
patient_count = np.array([32, 24, 28, 17, 13, 6, 3, 3, 2])
|
||||||
|
|
||||||
|
# "Remaining" = patients with >= that many visits (cumulative from the right)
|
||||||
|
remaining = np.array([patient_count[i:].sum() for i in range(len(patient_count))])
|
||||||
|
|
||||||
|
# --- Plot ---
|
||||||
|
fig, ax = plt.subplots(figsize=(12, 3))
|
||||||
|
|
||||||
|
y = 0.0 # all bubbles on one horizontal line
|
||||||
|
|
||||||
|
# Horizontal line
|
||||||
|
ax.hlines(y, visits.min() - 0.4, visits.max() + 0.4, color="#1f77b4", linewidth=3)
|
||||||
|
|
||||||
|
# Bubble sizes (scale as needed)
|
||||||
|
# (Matplotlib scatter uses area in points^2)
|
||||||
|
sizes = patient_count * 35 # tweak this multiplier if you want bigger/smaller bubbles
|
||||||
|
|
||||||
|
ax.scatter(visits, np.full_like(visits, y), s=sizes, color="#1f77b4", zorder=3)
|
||||||
|
|
||||||
|
# Title
|
||||||
|
#ax.set_title("Patient Journey by Visit Count", fontsize=14, pad=18)
|
||||||
|
|
||||||
|
# Top labels: "1 visits", "2 visits", ...
|
||||||
|
for x in visits:
|
||||||
|
label = f"{x} visit" if x == 1 else f"{x} visits"
|
||||||
|
ax.text(x, y + 0.18, label, ha="center", va="bottom", fontsize=10)
|
||||||
|
|
||||||
|
# Bottom labels: "X patients" and "Y remaining"
|
||||||
|
for x, pc, rem in zip(visits, patient_count, remaining):
|
||||||
|
ax.text(x, y - 0.20, f"{pc} patients", ha="center", va="top", fontsize=9)
|
||||||
|
ax.text(x, y - 0.32, f"{rem} remaining", ha="center", va="top", fontsize=9)
|
||||||
|
|
||||||
|
# Cosmetics: remove axes, keep spacing nice
|
||||||
|
ax.set_xlim(visits.min() - 0.6, visits.max() + 0.6)
|
||||||
|
ax.set_ylim(-0.5, 0.35)
|
||||||
|
ax.set_xticks([])
|
||||||
|
ax.set_yticks([])
|
||||||
|
for spine in ax.spines.values():
|
||||||
|
spine.set_visible(False)
|
||||||
|
|
||||||
|
plt.tight_layout()
|
||||||
|
plt.show()
|
||||||
|
plt.savefig("patient_journey.svg", format="svg", bbox_inches="tight")
|
||||||
|
##
|
||||||
|
|
||||||
|
|||||||
File diff suppressed because it is too large
Load Diff
Reference in New Issue
Block a user