1236 lines
45 KiB
Python
1236 lines
45 KiB
Python
# %% Explore Dist Plot
|
|
import pandas as pd
|
|
import json
|
|
import glob
|
|
import os
|
|
import re
|
|
import matplotlib.pyplot as plt
|
|
|
|
def plot_edss_distribution_per_iteration(json_dir_path):
|
|
# 1. Reuse your categorization logic
|
|
def categorize_edss(value):
|
|
if pd.isna(value): return 'Unknown'
|
|
elif value <= 1.0: return '0-1'
|
|
elif value <= 2.0: return '1-2'
|
|
elif value <= 3.0: return '2-3'
|
|
elif value <= 4.0: return '3-4'
|
|
elif value <= 5.0: return '4-5'
|
|
elif value <= 6.0: return '5-6'
|
|
elif value <= 7.0: return '6-7'
|
|
elif value <= 8.0: return '7-8'
|
|
elif value <= 9.0: return '8-9'
|
|
elif value <= 10.0: return '9-10'
|
|
else: return '10+'
|
|
|
|
# 2. Extract data from all files with Numerical Sorting
|
|
all_records = []
|
|
json_files = glob.glob(os.path.join(json_dir_path, "*.json"))
|
|
|
|
# Natural sort function to handle Iter 1, Iter 2 ... Iter 10
|
|
def natural_key(string_):
|
|
return [int(s) if s.isdigit() else s for s in re.split(r'(\d+)', string_)]
|
|
|
|
json_files.sort(key=natural_key)
|
|
|
|
for i, file_path in enumerate(json_files):
|
|
# We use the index + 1 for the label to ensure Iter 1 to Iter 10 order
|
|
iter_label = f"Iter {i+1}"
|
|
with open(file_path, 'r', encoding='utf-8') as f:
|
|
try:
|
|
data = json.load(f)
|
|
for entry in data:
|
|
if entry.get("success"):
|
|
val = entry["result"].get("EDSS")
|
|
all_records.append({
|
|
'Iteration': iter_label,
|
|
'Category': categorize_edss(val),
|
|
'Order': i # Used to maintain sort order in the table
|
|
})
|
|
except Exception as e:
|
|
print(f"Error reading {file_path}: {e}")
|
|
|
|
df = pd.DataFrame(all_records)
|
|
|
|
# 3. Create a Frequency Table (Crosstab)
|
|
# Pivot so iterations are on the X-axis
|
|
dist_table = pd.crosstab(df['Iteration'], df['Category'])
|
|
|
|
# Ensure the rows (Iterations) stay in the 1-10 order
|
|
iter_order = [f"Iter {i+1}" for i in range(len(json_files))]
|
|
dist_table = dist_table.reindex(iter_order)
|
|
|
|
# Ensure columns follow clinical order
|
|
fixed_labels = ['0-1', '1-2', '2-3', '3-4', '4-5', '5-6', '6-7', '7-8', '8-9', '9-10']
|
|
available_labels = [l for l in fixed_labels if l in dist_table.columns]
|
|
dist_table = dist_table[available_labels]
|
|
|
|
# 4. Plotting
|
|
ax = dist_table.plot(kind='bar', stacked=True, figsize=(14, 8), colormap='viridis', edgecolor='white')
|
|
|
|
plt.title('Distribution of Predicted EDSS Categories per Iteration', fontsize=15, pad=20)
|
|
plt.xlabel('JSON Iteration File', fontsize=12)
|
|
plt.ylabel('Number of Cases (Count)', fontsize=12)
|
|
plt.xticks(rotation=0)
|
|
|
|
# Move legend outside to the right
|
|
plt.legend(title="EDSS Category", bbox_to_anchor=(1.05, 1), loc='upper left')
|
|
|
|
# Add total count labels on top of bars
|
|
for i, (name, row) in enumerate(dist_table.iterrows()):
|
|
total = row.sum()
|
|
if total > 0:
|
|
plt.text(i, total + 2, f'Total: {int(total)}', ha='center', va='bottom', fontweight='bold')
|
|
|
|
plt.tight_layout()
|
|
plt.show()
|
|
|
|
return dist_table
|
|
# Usage:
|
|
counts_table = plot_edss_distribution_per_iteration('/home/shahin/Lab/Doktorarbeit/Barcelona/Data/iteration')
|
|
print(counts_table)
|
|
##
|
|
|
|
|
|
# %% Explore Table
|
|
import pandas as pd
|
|
import json
|
|
import glob
|
|
import os
|
|
import re
|
|
|
|
def generate_edss_distribution_csv(json_dir_path, output_filename='edss_distribution_summary.csv'):
|
|
# 1. Categorization logic
|
|
def categorize_edss(value):
|
|
if pd.isna(value): return 'Unknown'
|
|
elif value <= 1.0: return '0-1'
|
|
elif value <= 2.0: return '1-2'
|
|
elif value <= 3.0: return '2-3'
|
|
elif value <= 4.0: return '3-4'
|
|
elif value <= 5.0: return '4-5'
|
|
elif value <= 6.0: return '5-6'
|
|
elif value <= 7.0: return '6-7'
|
|
elif value <= 8.0: return '7-8'
|
|
elif value <= 9.0: return '8-9'
|
|
elif value <= 10.0: return '9-10'
|
|
else: return '10+'
|
|
|
|
# 2. Extract data from files with Natural Sorting
|
|
all_records = []
|
|
json_files = glob.glob(os.path.join(json_dir_path, "*.json"))
|
|
|
|
def natural_key(string_):
|
|
return [int(s) if s.isdigit() else s for s in re.split(r'(\d+)', string_)]
|
|
|
|
json_files.sort(key=natural_key)
|
|
|
|
for i, file_path in enumerate(json_files):
|
|
iter_label = f"Iter {i+1}"
|
|
with open(file_path, 'r', encoding='utf-8') as f:
|
|
try:
|
|
data = json.load(f)
|
|
for entry in data:
|
|
if entry.get("success"):
|
|
val = entry["result"].get("EDSS")
|
|
all_records.append({
|
|
'Iteration': iter_label,
|
|
'Category': categorize_edss(val)
|
|
})
|
|
except Exception as e:
|
|
print(f"Error reading {file_path}: {e}")
|
|
|
|
df = pd.DataFrame(all_records)
|
|
|
|
# 3. Create Frequency Table (Crosstab)
|
|
dist_table = pd.crosstab(df['Iteration'], df['Category'])
|
|
|
|
# 4. Reindex Rows (Numerical order) and Columns (Clinical order)
|
|
iter_order = [f"Iter {i+1}" for i in range(len(json_files))]
|
|
dist_table = dist_table.reindex(iter_order)
|
|
|
|
fixed_labels = ['0-1', '1-2', '2-3', '3-4', '4-5', '5-6', '6-7', '7-8', '8-9', '9-10']
|
|
available_labels = [l for l in fixed_labels if l in dist_table.columns]
|
|
dist_table = dist_table[available_labels]
|
|
|
|
# Fill missing categories with 0 and convert to integers
|
|
dist_table = dist_table.fillna(0).astype(int)
|
|
|
|
# 5. Add "Total" row at the end
|
|
# This sums the counts for each category across all iterations
|
|
dist_table.loc['Total Sum'] = dist_table.sum()
|
|
|
|
# 6. Save to CSV
|
|
dist_table.to_csv(output_filename)
|
|
print(f"Table successfully saved to: {output_filename}")
|
|
|
|
return dist_table
|
|
|
|
# Usage:
|
|
final_table = generate_edss_distribution_csv('/home/shahin/Lab/Doktorarbeit/Barcelona/Data/iteration')
|
|
print(final_table)
|
|
##
|
|
|
|
# %% EDSS Confusion Matrix
|
|
import pandas as pd
|
|
import numpy as np
|
|
import json
|
|
import glob
|
|
import os
|
|
import matplotlib.pyplot as plt
|
|
from sklearn.metrics import confusion_matrix, ConfusionMatrixDisplay
|
|
|
|
def categorize_edss(value):
|
|
if pd.isna(value):
|
|
return np.nan
|
|
elif value <= 1.0:
|
|
return '0-1'
|
|
elif value <= 2.0:
|
|
return '1-2'
|
|
elif value <= 3.0:
|
|
return '2-3'
|
|
elif value <= 4.0:
|
|
return '3-4'
|
|
elif value <= 5.0:
|
|
return '4-5'
|
|
elif value <= 6.0:
|
|
return '5-6'
|
|
elif value <= 7.0:
|
|
return '6-7'
|
|
elif value <= 8.0:
|
|
return '7-8'
|
|
elif value <= 9.0:
|
|
return '8-9'
|
|
elif value <= 10.0:
|
|
return '9-10'
|
|
else:
|
|
return '10+'
|
|
|
|
def plot_categorized_edss(json_dir_path, ground_truth_path):
|
|
# 1. Load Ground Truth
|
|
df_gt = pd.read_csv(ground_truth_path, sep=';')
|
|
df_gt['unique_id'] = df_gt['unique_id'].astype(str)
|
|
df_gt['MedDatum'] = df_gt['MedDatum'].astype(str)
|
|
df_gt['EDSS'] = pd.to_numeric(df_gt['EDSS'], errors='coerce')
|
|
|
|
# 2. Iterate through JSON files
|
|
all_preds = []
|
|
json_pattern = os.path.join(json_dir_path, "*.json")
|
|
for file_path in glob.glob(json_pattern):
|
|
with open(file_path, 'r', encoding='utf-8') as f:
|
|
try:
|
|
data = json.load(f)
|
|
for entry in data:
|
|
if entry.get("success") and "result" in entry:
|
|
res = entry["result"]
|
|
all_preds.append({
|
|
'unique_id': str(res.get('unique_id')),
|
|
'MedDatum': str(res.get('MedDatum')),
|
|
'edss_pred': res.get('EDSS')
|
|
})
|
|
except Exception as e:
|
|
print(f"Error reading {file_path}: {e}")
|
|
|
|
df_pred = pd.DataFrame(all_preds)
|
|
df_pred['edss_pred'] = pd.to_numeric(df_pred['edss_pred'], errors='coerce')
|
|
|
|
# 3. Merge and Categorize
|
|
# Clean keys to ensure 100% match rate
|
|
for df in [df_gt, df_pred]:
|
|
df['unique_id'] = df['unique_id'].astype(str).str.strip()
|
|
df['MedDatum'] = df['MedDatum'].astype(str).str.strip()
|
|
|
|
df_merged = pd.merge(
|
|
df_gt[['unique_id', 'MedDatum', 'EDSS']],
|
|
df_pred,
|
|
on=['unique_id', 'MedDatum'],
|
|
how='inner'
|
|
)
|
|
|
|
df_merged = df_merged.dropna(subset=['EDSS', 'edss_pred'])
|
|
|
|
# --- ADDED THESE LINES TO FIX THE NAMEERROR ---
|
|
y_true = df_merged['EDSS'].apply(categorize_edss)
|
|
y_pred = df_merged['edss_pred'].apply(categorize_edss)
|
|
# ----------------------------------------------
|
|
|
|
print(f"Verification: Total matches in Confusion Matrix: {len(df_merged)}")
|
|
|
|
# 4. Define fixed labels to handle data gaps
|
|
fixed_labels = ['0-1', '1-2', '2-3', '3-4', '4-5', '5-6', '6-7', '7-8', '8-9', '9-10']
|
|
|
|
# 5. Generate Confusion Matrix
|
|
cm = confusion_matrix(y_true, y_pred, labels=fixed_labels)
|
|
|
|
# 6. Plotting
|
|
fig, ax = plt.subplots(figsize=(10, 8))
|
|
disp = ConfusionMatrixDisplay(confusion_matrix=cm, display_labels=fixed_labels)
|
|
|
|
# Plotting (y_axis is Ground Truth, x_axis is LLM Prediction)
|
|
disp.plot(cmap=plt.cm.Blues, values_format='d', ax=ax)
|
|
|
|
plt.title('Categorized EDSS: Ground Truth vs LLM Prediction')
|
|
plt.ylabel('Ground Truth EDSS')
|
|
plt.xlabel('LLM Prediction')
|
|
plt.show()
|
|
##
|
|
|
|
# %% Confusion Matrix adjusted
|
|
import pandas as pd
|
|
import numpy as np
|
|
import json
|
|
import glob
|
|
import os
|
|
import matplotlib.pyplot as plt
|
|
from sklearn.metrics import confusion_matrix, ConfusionMatrixDisplay
|
|
|
|
def categorize_edss(value):
|
|
"""Bins EDSS values into clinical categories."""
|
|
if pd.isna(value):
|
|
return np.nan
|
|
elif value <= 1.0: return '0-1'
|
|
elif value <= 2.0: return '1-2'
|
|
elif value <= 3.0: return '2-3'
|
|
elif value <= 4.0: return '3-4'
|
|
elif value <= 5.0: return '4-5'
|
|
elif value <= 6.0: return '5-6'
|
|
elif value <= 7.0: return '6-7'
|
|
elif value <= 8.0: return '7-8'
|
|
elif value <= 9.0: return '8-9'
|
|
elif value <= 10.0: return '9-10'
|
|
else: return '10+'
|
|
|
|
def plot_categorized_edss(json_dir_path, ground_truth_path):
|
|
# 1. Load Ground Truth with Normalization
|
|
df_gt = pd.read_csv(ground_truth_path, sep=';')
|
|
# Standardize keys to ensure 1:N matching works
|
|
df_gt['unique_id'] = df_gt['unique_id'].astype(str).str.strip().str.lower()
|
|
df_gt['MedDatum'] = df_gt['MedDatum'].astype(str).str.strip().str.lower()
|
|
df_gt['EDSS'] = pd.to_numeric(df_gt['EDSS'], errors='coerce')
|
|
|
|
# 2. Load All Predictions from JSONs
|
|
all_preds = []
|
|
json_files = glob.glob(os.path.join(json_dir_path, "*.json"))
|
|
|
|
for file_path in json_files:
|
|
with open(file_path, 'r', encoding='utf-8') as f:
|
|
try:
|
|
data = json.load(f)
|
|
for entry in data:
|
|
# We only take 'success': true entries
|
|
if entry.get("success") and "result" in entry:
|
|
res = entry["result"]
|
|
all_preds.append({
|
|
'unique_id': str(res.get('unique_id')).strip().lower(),
|
|
'MedDatum': str(res.get('MedDatum')).strip().lower(),
|
|
'edss_pred': res.get('EDSS')
|
|
})
|
|
except Exception as e:
|
|
print(f"Error reading {file_path}: {e}")
|
|
|
|
df_pred = pd.DataFrame(all_preds)
|
|
df_pred['edss_pred'] = pd.to_numeric(df_pred['edss_pred'], errors='coerce')
|
|
|
|
# 3. Merge (This should give you ~3934 rows based on your audit)
|
|
df_merged = pd.merge(
|
|
df_gt[['unique_id', 'MedDatum', 'EDSS']],
|
|
df_pred,
|
|
on=['unique_id', 'MedDatum'],
|
|
how='inner'
|
|
)
|
|
|
|
# --- THE BIG REVEAL: Count the NaNs ---
|
|
nan_in_gt = df_merged['EDSS'].isna().sum()
|
|
nan_in_pred = df_merged['edss_pred'].isna().sum()
|
|
|
|
print("-" * 40)
|
|
print(f"TOTAL MERGED ROWS: {len(df_merged)}")
|
|
print(f"Rows with missing Ground Truth EDSS: {nan_in_gt}")
|
|
print(f"Rows with missing Prediction EDSS: {nan_in_pred}")
|
|
print("-" * 40)
|
|
|
|
# Now drop rows that have NO values in either side for the matrix
|
|
df_final = df_merged.dropna(subset=['EDSS', 'edss_pred']).copy()
|
|
print(f"FINAL ROWS FOR CONFUSION MATRIX: {len(df_final)}")
|
|
print("-" * 40)
|
|
|
|
# 4. Categorize for the Matrix
|
|
y_true = df_final['EDSS'].apply(categorize_edss)
|
|
y_pred = df_final['edss_pred'].apply(categorize_edss)
|
|
|
|
fixed_labels = ['0-1', '1-2', '2-3', '3-4', '4-5', '5-6', '6-7', '7-8', '8-9', '9-10']
|
|
|
|
# 5. Generate and Print Raw Matrix
|
|
cm = confusion_matrix(y_true, y_pred, labels=fixed_labels)
|
|
|
|
# Print the Raw Matrix to terminal
|
|
cm_df = pd.DataFrame(cm, index=[f"True_{l}" for l in fixed_labels],
|
|
columns=[f"Pred_{l}" for l in fixed_labels])
|
|
print("\nRAW CONFUSION MATRIX (Rows=True, Cols=Pred):")
|
|
print(cm_df)
|
|
|
|
# 6. Plotting
|
|
fig, ax = plt.subplots(figsize=(12, 10))
|
|
disp = ConfusionMatrixDisplay(confusion_matrix=cm, display_labels=fixed_labels)
|
|
|
|
# Values_format='d' ensures we see whole numbers, not scientific notation
|
|
disp.plot(cmap=plt.cm.Blues, values_format='d', ax=ax)
|
|
|
|
plt.title(f'EDSS Confusion Matrix\n(n={len(df_final)} iterations across ~400 cases)', fontsize=14)
|
|
plt.ylabel('Ground Truth (Clinician)')
|
|
plt.xlabel('LLM Prediction')
|
|
plt.xticks(rotation=45)
|
|
plt.tight_layout()
|
|
plt.show()
|
|
|
|
##
|
|
# %% Subcategories
|
|
|
|
import pandas as pd
|
|
import numpy as np
|
|
import json
|
|
import glob
|
|
import os
|
|
import matplotlib.pyplot as plt
|
|
|
|
def plot_subcategory_analysis(json_dir_path, ground_truth_path):
|
|
# 1. Column Mapping (JSON Key : CSV Column)
|
|
mapping = {
|
|
"VISUAL_OPTIC_FUNCTIONS": "Sehvermögen",
|
|
"BRAINSTEM_FUNCTIONS": "Hirnstamm",
|
|
"PYRAMIDAL_FUNCTIONS": "Pyramidalmotorik",
|
|
"CEREBELLAR_FUNCTIONS": "Cerebellum",
|
|
"SENSORY_FUNCTIONS": "Sensibiliät",
|
|
"BOWEL_AND_BLADDER_FUNCTIONS": "Blasen-_und_Mastdarmfunktion",
|
|
"CEREBRAL_FUNCTIONS": "Cerebrale_Funktion",
|
|
"AMBULATION": "Ambulation"
|
|
}
|
|
|
|
# 2. Load Ground Truth
|
|
df_gt = pd.read_csv(ground_truth_path, sep=';')
|
|
df_gt['unique_id'] = df_gt['unique_id'].astype(str)
|
|
df_gt['MedDatum'] = df_gt['MedDatum'].astype(str)
|
|
|
|
# 3. Load Predictions including Subcategories
|
|
all_preds = []
|
|
for file_path in glob.glob(os.path.join(json_dir_path, "*.json")):
|
|
with open(file_path, 'r', encoding='utf-8') as f:
|
|
data = json.load(f)
|
|
for entry in data:
|
|
if entry.get("success"):
|
|
res = entry["result"]
|
|
row = {
|
|
'unique_id': str(res.get('unique_id')),
|
|
'MedDatum': str(res.get('MedDatum'))
|
|
}
|
|
# Add subcategory scores
|
|
for json_key in mapping.keys():
|
|
row[json_key] = res.get('subcategories', {}).get(json_key)
|
|
all_preds.append(row)
|
|
|
|
df_pred = pd.DataFrame(all_preds)
|
|
|
|
# 4. Merge
|
|
df_merged = pd.merge(df_gt, df_pred, on=['unique_id', 'MedDatum'], suffixes=('_gt', '_llm'))
|
|
|
|
# 5. Calculate Metrics
|
|
results = []
|
|
for json_key, csv_col in mapping.items():
|
|
# Ensure numeric
|
|
true_vals = pd.to_numeric(df_merged[csv_col], errors='coerce')
|
|
pred_vals = pd.to_numeric(df_merged[json_key], errors='coerce')
|
|
|
|
# Drop NaNs for this specific subcategory
|
|
mask = true_vals.notna() & pred_vals.notna()
|
|
y_t = true_vals[mask]
|
|
y_p = pred_vals[mask]
|
|
|
|
if len(y_t) > 0:
|
|
accuracy = (y_t == y_p).mean() * 100
|
|
mae = np.abs(y_t - y_p).mean() # Mean Absolute Error (Deviation)
|
|
results.append({
|
|
'Subcategory': csv_col,
|
|
'Accuracy': accuracy,
|
|
'Deviation': mae
|
|
})
|
|
|
|
stats_df = pd.DataFrame(results).sort_values('Accuracy', ascending=False)
|
|
|
|
# 6. Plotting
|
|
fig, ax1 = plt.subplots(figsize=(14, 7))
|
|
|
|
# Bar chart for Accuracy
|
|
bars = ax1.bar(stats_df['Subcategory'], stats_df['Accuracy'],
|
|
color='#3498db', alpha=0.8, label='Accuracy (%)')
|
|
ax1.set_ylabel('Accuracy (%)', color='#2980b9', fontsize=12, fontweight='bold')
|
|
ax1.set_ylim(0, 115) # Extra head room for labels
|
|
ax1.grid(axis='y', linestyle='--', alpha=0.7)
|
|
|
|
# Rotate labels
|
|
plt.xticks(rotation=30, ha='right', fontsize=10)
|
|
|
|
# Line chart for Deviation
|
|
ax2 = ax1.twinx()
|
|
ax2.plot(stats_df['Subcategory'], stats_df['Deviation'],
|
|
color='#e74c3c', marker='o', linewidth=2.5, markersize=8,
|
|
label='Mean Abs. Deviation (Score Points)')
|
|
ax2.set_ylabel('Mean Absolute Deviation', color='#c0392b', fontsize=12, fontweight='bold')
|
|
|
|
# Adjust ax2 limit to avoid overlap with accuracy text
|
|
ax2.set_ylim(0, max(stats_df['Deviation']) * 1.5 if not stats_df['Deviation'].empty else 1)
|
|
|
|
# plt.title('Subcategory Performance: Accuracy vs. Mean Deviation', fontsize=14, pad=20)
|
|
|
|
# --- THE FIX: Better Legend Placement ---
|
|
# Combine legends from both axes and place them above the plot
|
|
lines1, labels1 = ax1.get_legend_handles_labels()
|
|
lines2, labels2 = ax2.get_legend_handles_labels()
|
|
ax1.legend(lines1 + lines2, labels1 + labels2,
|
|
loc='upper center', bbox_to_anchor=(0.5, 1.12),
|
|
ncol=2, frameon=False, fontsize=11)
|
|
|
|
# Add percentage labels on top of bars
|
|
for bar in bars:
|
|
height = bar.get_height()
|
|
ax1.annotate(f'{height:.1f}%',
|
|
xy=(bar.get_x() + bar.get_width() / 2, height),
|
|
xytext=(0, 5), textcoords="offset points",
|
|
ha='center', va='bottom', fontweight='bold', color='#2c3e50')
|
|
|
|
plt.tight_layout()
|
|
plt.show()
|
|
##
|
|
|
|
# %% Certainty
|
|
import pandas as pd
|
|
import numpy as np
|
|
import json
|
|
import glob
|
|
import os
|
|
import matplotlib.pyplot as plt
|
|
|
|
def categorize_edss(value):
|
|
if pd.isna(value): return np.nan
|
|
elif value <= 1.0: return '0-1'
|
|
elif value <= 2.0: return '1-2'
|
|
elif value <= 3.0: return '2-3'
|
|
elif value <= 4.0: return '3-4'
|
|
elif value <= 5.0: return '4-5'
|
|
elif value <= 6.0: return '5-6'
|
|
elif value <= 7.0: return '6-7'
|
|
elif value <= 8.0: return '7-8'
|
|
elif value <= 9.0: return '8-9'
|
|
elif value <= 10.0: return '9-10'
|
|
else: return '10+'
|
|
|
|
def plot_certainty_vs_accuracy_by_category(json_dir_path, ground_truth_path):
|
|
# 1. Data Loading & Merging
|
|
df_gt = pd.read_csv(ground_truth_path, sep=';')
|
|
df_gt['unique_id'] = df_gt['unique_id'].astype(str)
|
|
df_gt['MedDatum'] = df_gt['MedDatum'].astype(str)
|
|
df_gt['EDSS_true'] = pd.to_numeric(df_gt['EDSS'], errors='coerce')
|
|
|
|
all_preds = []
|
|
for file_path in glob.glob(os.path.join(json_dir_path, "*.json")):
|
|
with open(file_path, 'r', encoding='utf-8') as f:
|
|
data = json.load(f)
|
|
for entry in data:
|
|
if entry.get("success"):
|
|
res = entry["result"]
|
|
all_preds.append({
|
|
'unique_id': str(res.get('unique_id')),
|
|
'MedDatum': str(res.get('MedDatum')),
|
|
'EDSS_pred': res.get('EDSS'),
|
|
'certainty': res.get('certainty_percent')
|
|
})
|
|
|
|
df_pred = pd.DataFrame(all_preds)
|
|
df_pred['EDSS_pred'] = pd.to_numeric(df_pred['EDSS_pred'], errors='coerce')
|
|
|
|
df = pd.merge(df_gt[['unique_id', 'MedDatum', 'EDSS_true']],
|
|
df_pred, on=['unique_id', 'MedDatum']).dropna()
|
|
|
|
# 2. Process Metrics
|
|
df['gt_category'] = df['EDSS_true'].apply(categorize_edss)
|
|
df['is_correct'] = (df['EDSS_true'].round(1) == df['EDSS_pred'].round(1))
|
|
|
|
fixed_labels = ['0-1', '1-2', '2-3', '3-4', '4-5', '5-6', '6-7', '7-8', '8-9', '9-10']
|
|
|
|
# Calculate Mean Certainty and Mean Accuracy per category
|
|
stats = df.groupby('gt_category').agg({
|
|
'is_correct': 'mean',
|
|
'certainty': 'mean',
|
|
'unique_id': 'count'
|
|
}).reindex(fixed_labels)
|
|
|
|
stats['accuracy_percent'] = stats['is_correct'] * 100
|
|
stats = stats.fillna(0)
|
|
|
|
# 3. Plotting
|
|
x = np.arange(len(fixed_labels))
|
|
width = 0.35 # Width of the bars
|
|
|
|
fig, ax = plt.subplots(figsize=(14, 8))
|
|
|
|
# Plotting both bars side-by-side
|
|
rects1 = ax.bar(x - width/2, stats['accuracy_percent'], width,
|
|
label='Actual Accuracy (%)', color='#2ecc71', alpha=0.8)
|
|
rects2 = ax.bar(x + width/2, stats['certainty'], width,
|
|
label='LLM Avg. Certainty (%)', color='#e67e22', alpha=0.8)
|
|
|
|
# Add text labels, titles and custom x-axis tick labels, etc.
|
|
ax.set_ylabel('Percentage (%)', fontsize=12)
|
|
ax.set_xlabel('Ground Truth EDSS Category', fontsize=12)
|
|
# ax.set_title('Comparison: LLM Confidence (Certainty) vs. Real Accuracy per EDSS Range', fontsize=15, pad=25)
|
|
ax.set_xticks(x)
|
|
ax.set_xticklabels(fixed_labels)
|
|
ax.set_ylim(0, 115)
|
|
ax.legend(loc='upper center', bbox_to_anchor=(0.5, 1.08), ncol=2, frameon=False)
|
|
ax.grid(axis='y', linestyle=':', alpha=0.5)
|
|
|
|
# Helper function to label bar heights
|
|
def autolabel(rects):
|
|
for rect in rects:
|
|
height = rect.get_height()
|
|
if height > 0:
|
|
ax.annotate(f'{height:.0f}%',
|
|
xy=(rect.get_x() + rect.get_width() / 2, height),
|
|
xytext=(0, 3), textcoords="offset points",
|
|
ha='center', va='bottom', fontsize=9, fontweight='bold')
|
|
|
|
autolabel(rects1)
|
|
autolabel(rects2)
|
|
|
|
# Add sample size (n) at the bottom
|
|
for i, count in enumerate(stats['unique_id']):
|
|
ax.text(i, 2, f'n={int(count)}', ha='center', va='bottom', fontsize=10, color='white', fontweight='bold')
|
|
|
|
plt.tight_layout()
|
|
plt.show()
|
|
|
|
##
|
|
|
|
|
|
|
|
# %% Boxplot
|
|
import pandas as pd
|
|
import numpy as np
|
|
import json
|
|
import glob
|
|
import os
|
|
import re
|
|
import matplotlib.pyplot as plt
|
|
from matplotlib.lines import Line2D
|
|
from matplotlib.patches import Patch
|
|
|
|
def natural_key(string_):
|
|
return [int(s) if s.isdigit() else s for s in re.split(r'(\d+)', string_)]
|
|
|
|
def plot_edss_boxplot(json_dir_path, ground_truth_path):
|
|
# 1. Load Ground Truth
|
|
df_gt = pd.read_csv(ground_truth_path, sep=';')
|
|
gt_values = pd.to_numeric(df_gt['EDSS'], errors='coerce').dropna().tolist()
|
|
|
|
# 2. Load Iterations
|
|
json_files = glob.glob(os.path.join(json_dir_path, "*.json"))
|
|
json_files.sort(key=natural_key)
|
|
|
|
plot_data = [gt_values]
|
|
labels = ['Ground Truth']
|
|
|
|
for i, file_path in enumerate(json_files):
|
|
iteration_values = []
|
|
with open(file_path, 'r', encoding='utf-8') as f:
|
|
data = json.load(f)
|
|
for entry in data:
|
|
if entry.get("success"):
|
|
val = entry["result"].get("EDSS")
|
|
if val is not None:
|
|
iteration_values.append(float(val))
|
|
plot_data.append(iteration_values)
|
|
labels.append(f"Iter {i+1}")
|
|
|
|
# 3. Plotting Configuration
|
|
plt.figure(figsize=(14, 8))
|
|
|
|
# Define colors
|
|
gt_color = '#ff9999' # Soft Red
|
|
iter_color = '#66b3ff' # Soft Blue
|
|
|
|
# Create the boxplot
|
|
bplot = plt.boxplot(plot_data, labels=labels, patch_artist=True,
|
|
notch=False,
|
|
medianprops={'color': 'black', 'linewidth': 2},
|
|
flierprops={'marker': 'o', 'markerfacecolor': 'gray', 'markersize': 5, 'alpha': 0.5},
|
|
showmeans=True,
|
|
meanprops={"marker":"D", "markerfacecolor":"white", "markeredgecolor":"black", "markersize": 6})
|
|
|
|
# 4. Fill boxes with colors
|
|
colors = [gt_color] + [iter_color] * (len(plot_data) - 1)
|
|
for patch, color in zip(bplot['boxes'], colors):
|
|
patch.set_facecolor(color)
|
|
|
|
# 5. CONSTRUCT THE COMPLETE LEGEND
|
|
legend_elements = [
|
|
Patch(facecolor=gt_color, edgecolor='black', label='Ground Truth'),
|
|
Patch(facecolor=iter_color, edgecolor='black', label='LLM Iterations (1-10)'),
|
|
Line2D([0], [0], color='black', lw=2, label='Median'),
|
|
Line2D([0], [0], marker='D', color='w', label='Mean Score',
|
|
markerfacecolor='white', markeredgecolor='black', markersize=8),
|
|
Line2D([0], [0], marker='o', color='w', label='Outliers',
|
|
markerfacecolor='gray', markersize=6, alpha=0.5)
|
|
]
|
|
|
|
plt.legend(handles=legend_elements, loc='upper right', frameon=True, shadow=True, title="Legend")
|
|
|
|
# Formatting
|
|
plt.title('Distribution of EDSS Scores: Ground Truth vs. 10 LLM Iterations', fontsize=16, pad=20)
|
|
plt.ylabel('EDSS Score (0-10)', fontsize=12)
|
|
plt.xlabel('Data Source', fontsize=12)
|
|
plt.grid(axis='y', linestyle='--', alpha=0.4)
|
|
plt.ylim(-0.5, 10.5)
|
|
plt.xticks(rotation=45)
|
|
|
|
plt.tight_layout()
|
|
plt.show()
|
|
##
|
|
|
|
# %% Audit
|
|
|
|
|
|
import pandas as pd
|
|
import numpy as np
|
|
import json
|
|
import glob
|
|
import os
|
|
|
|
def audit_matches(json_dir_path, ground_truth_path):
|
|
# 1. Load GT
|
|
df_gt = pd.read_csv(ground_truth_path, sep=';')
|
|
|
|
# 2. Advanced Normalization
|
|
def clean_series(s):
|
|
return s.astype(str).str.strip().str.lower()
|
|
|
|
df_gt['unique_id'] = clean_series(df_gt['unique_id'])
|
|
df_gt['MedDatum'] = clean_series(df_gt['MedDatum'])
|
|
|
|
# 3. Load Predictions
|
|
all_preds = []
|
|
for file_path in glob.glob(os.path.join(json_dir_path, "*.json")):
|
|
with open(file_path, 'r', encoding='utf-8') as f:
|
|
data = json.load(f)
|
|
file_name = os.path.basename(file_path)
|
|
for entry in data:
|
|
if entry.get("success"):
|
|
res = entry["result"]
|
|
all_preds.append({
|
|
'unique_id': str(res.get('unique_id')).strip().lower(),
|
|
'MedDatum': str(res.get('MedDatum')).strip().lower(),
|
|
'file': file_name
|
|
})
|
|
|
|
df_pred = pd.DataFrame(all_preds)
|
|
|
|
# 4. Find the "Ghost" entries (In JSON but not in GT)
|
|
# Create a 'key' column for easy comparison
|
|
df_gt['key'] = df_gt['unique_id'] + "_" + df_gt['MedDatum']
|
|
df_pred['key'] = df_pred['unique_id'] + "_" + df_pred['MedDatum']
|
|
|
|
gt_keys = set(df_gt['key'])
|
|
df_pred['is_matched'] = df_pred['key'].isin(gt_keys)
|
|
|
|
unmatched_summary = df_pred[df_pred['is_matched'] == False]
|
|
|
|
print("--- AUDIT RESULTS ---")
|
|
print(f"Total rows in JSON: {len(df_pred)}")
|
|
print(f"Rows that matched GT: {df_pred['is_matched'].sum()}")
|
|
print(f"Rows that FAILED to match: {len(unmatched_summary)}")
|
|
|
|
if not unmatched_summary.empty:
|
|
print("\nFirst 10 Unmatched Entries (check these against your CSV):")
|
|
print(unmatched_summary[['unique_id', 'MedDatum', 'file']].head(10))
|
|
|
|
# Breakdown by file - see if specific JSON files are broken
|
|
print("\nFailure count per JSON file:")
|
|
print(unmatched_summary['file'].value_counts())
|
|
|
|
|
|
##
|
|
|
|
|
|
|
|
|
|
# %% Cinfidence accuracy correlation
|
|
|
|
import pandas as pd
|
|
import numpy as np
|
|
import json
|
|
import glob
|
|
import os
|
|
import matplotlib.pyplot as plt
|
|
|
|
def categorize_edss(value):
|
|
if pd.isna(value): return np.nan
|
|
elif value <= 1.0: return '0-1'
|
|
elif value <= 2.0: return '1-2'
|
|
elif value <= 3.0: return '2-3'
|
|
elif value <= 4.0: return '3-4'
|
|
elif value <= 5.0: return '4-5'
|
|
elif value <= 6.0: return '6-7'
|
|
elif value <= 7.0: return '7-8'
|
|
elif value <= 8.0: return '8-9'
|
|
elif value <= 9.0: return '9-10'
|
|
else: return '10+'
|
|
|
|
def plot_binned_calibration(json_dir_path, ground_truth_path):
|
|
# 1. Load Ground Truth
|
|
df_gt = pd.read_csv(ground_truth_path, sep=';')
|
|
df_gt['unique_id'] = df_gt['unique_id'].astype(str).str.strip().str.lower()
|
|
df_gt['MedDatum'] = df_gt['MedDatum'].astype(str).str.strip().str.lower()
|
|
df_gt['gt_cat'] = pd.to_numeric(df_gt['EDSS'], errors='coerce').apply(categorize_edss)
|
|
|
|
# 2. Load Predictions
|
|
all_preds = []
|
|
for file_path in glob.glob(os.path.join(json_dir_path, "*.json")):
|
|
with open(file_path, 'r', encoding='utf-8') as f:
|
|
data = json.load(f)
|
|
for entry in data:
|
|
if entry.get("success"):
|
|
res = entry["result"]
|
|
all_preds.append({
|
|
'unique_id': str(res.get('unique_id')).strip().lower(),
|
|
'MedDatum': str(res.get('MedDatum')).strip().lower(),
|
|
'pred_cat': categorize_edss(res.get('EDSS')),
|
|
'confidence': res.get('certainty_percent')
|
|
})
|
|
|
|
df_pred = pd.DataFrame(all_preds)
|
|
df_merged = pd.merge(df_pred, df_gt[['unique_id', 'MedDatum', 'gt_cat']],
|
|
on=['unique_id', 'MedDatum'], how='inner')
|
|
|
|
# Define correctness
|
|
df_merged['is_correct'] = (df_merged['pred_cat'] == df_merged['gt_cat']).astype(int)
|
|
|
|
# 3. Create Confidence Bins (e.g., 0-60, 60-70, 70-80, 80-90, 90-100)
|
|
bins = [0, 60, 70, 80, 90, 100]
|
|
labels = ['<60%', '60-70%', '70-80%', '80-90%', '90-100%']
|
|
df_merged['conf_bin'] = pd.cut(df_merged['confidence'], bins=bins, labels=labels)
|
|
|
|
# Calculate average accuracy per bin
|
|
calibration_stats = df_merged.groupby('conf_bin')['is_correct'].agg(['mean', 'count']).reset_index()
|
|
|
|
# 4. Plotting
|
|
plt.figure(figsize=(10, 6))
|
|
|
|
# Bar chart for actual accuracy
|
|
bars = plt.bar(calibration_stats['conf_bin'], calibration_stats['mean'],
|
|
color='skyblue', edgecolor='navy', alpha=0.7, label='Actual Accuracy')
|
|
|
|
# Add the "Perfect Calibration" line
|
|
# (If confidence is 95%, accuracy should be 0.95)
|
|
expected_x = np.arange(len(labels))
|
|
expected_y = [0.3, 0.65, 0.75, 0.85, 0.95] # Midpoints of the bins for visual reference
|
|
plt.plot(expected_x, expected_y, color='red', marker='o', linestyle='--',
|
|
linewidth=2, label='Perfect Calibration (Theoretical)')
|
|
|
|
# 5. Add text labels on top of bars to show sample size (how many cases in that bin)
|
|
for i, bar in enumerate(bars):
|
|
yval = bar.get_height()
|
|
count = calibration_stats.loc[i, 'count']
|
|
plt.text(bar.get_x() + bar.get_width()/2, yval + 0.02,
|
|
f'Acc: {yval:.1%}\n(n={count})', ha='center', va='bottom', fontsize=9)
|
|
|
|
# Legend and Labels
|
|
plt.title('Model Calibration: Does Confidence Match Accuracy?', fontsize=14, pad=15)
|
|
plt.xlabel('LLM Confidence Score Bin', fontsize=12)
|
|
plt.ylabel('Actual Accuracy (Correct Category %)', fontsize=12)
|
|
plt.ylim(0, 1.1)
|
|
plt.grid(axis='y', linestyle=':', alpha=0.5)
|
|
|
|
# Adding a clean, informative legend
|
|
plt.legend(loc='upper left', frameon=True, shadow=True)
|
|
|
|
plt.tight_layout()
|
|
plt.show()
|
|
##
|
|
|
|
|
|
|
|
# %% Confidence comparison
|
|
|
|
import pandas as pd
|
|
import numpy as np
|
|
import json
|
|
import glob
|
|
import os
|
|
import matplotlib.pyplot as plt
|
|
import seaborn as sns
|
|
from matplotlib.lines import Line2D
|
|
from matplotlib.patches import Patch
|
|
|
|
def plot_edss_confidence_comparison(json_dir_path, ground_truth_path):
|
|
# 1. Load Ground Truth
|
|
df_gt = pd.read_csv(ground_truth_path, sep=';')
|
|
df_gt['unique_id'] = df_gt['unique_id'].astype(str).str.strip().str.lower()
|
|
df_gt['MedDatum'] = df_gt['MedDatum'].astype(str).str.strip().str.lower()
|
|
df_gt['EDSS_gt'] = pd.to_numeric(df_gt['EDSS'], errors='coerce')
|
|
|
|
# 2. Load Predictions from all JSONs
|
|
all_preds = []
|
|
for file_path in glob.glob(os.path.join(json_dir_path, "*.json")):
|
|
with open(file_path, 'r', encoding='utf-8') as f:
|
|
try:
|
|
data = json.load(f)
|
|
for entry in data:
|
|
if entry.get("success"):
|
|
res = entry["result"]
|
|
all_preds.append({
|
|
'unique_id': str(res.get('unique_id')).strip().lower(),
|
|
'MedDatum': str(res.get('MedDatum')).strip().lower(),
|
|
'EDSS_pred': pd.to_numeric(res.get('EDSS'), errors='coerce'),
|
|
'confidence': pd.to_numeric(res.get('certainty_percent'), errors='coerce')
|
|
})
|
|
except Exception as e:
|
|
print(f"Skipping {file_path}: {e}")
|
|
|
|
df_pred = pd.DataFrame(all_preds)
|
|
|
|
# 3. Merge and Clean
|
|
df_merged = pd.merge(df_pred, df_gt[['unique_id', 'MedDatum', 'EDSS_gt']],
|
|
on=['unique_id', 'MedDatum'], how='inner')
|
|
df_plot = df_merged.dropna(subset=['EDSS_pred', 'EDSS_gt', 'confidence']).copy()
|
|
|
|
# 4. Bin Confidence (X-Axis Categories)
|
|
# We group confidence into bins to create a readable boxplot
|
|
bins = [0, 60, 70, 80, 90, 100]
|
|
labels = ['<60%', '60-70%', '70-80%', '80-90%', '90-100%']
|
|
df_plot['conf_bin'] = pd.cut(df_plot['confidence'], bins=bins, labels=labels)
|
|
|
|
# 5. Plotting
|
|
plt.figure(figsize=(14, 8))
|
|
|
|
# A. Boxplot: Shows the distribution of LLM PREDICTIONS
|
|
sns.boxplot(data=df_plot, x='conf_bin', y='EDSS_pred',
|
|
color='#3498db', width=0.5, showfliers=False,
|
|
boxprops=dict(alpha=0.4, edgecolor='navy'))
|
|
|
|
# B. Stripplot (Dots): Shows individual GROUND TRUTH scores
|
|
# We add jitter so dots don't hide each other
|
|
sns.stripplot(data=df_plot, x='conf_bin', y='EDSS_gt',
|
|
color='#e74c3c', alpha=0.4, jitter=0.2, size=5)
|
|
|
|
# 6. Create a CLEAR Legend
|
|
legend_elements = [
|
|
Patch(facecolor='#3498db', edgecolor='navy', alpha=0.4,
|
|
label='LLM Predictions (Box = Distribution)'),
|
|
Line2D([0], [0], marker='o', color='w', label='Ground Truth (Dots = Clinician Scores)',
|
|
markerfacecolor='#e74c3c', markersize=8, alpha=0.6),
|
|
Line2D([0], [0], color='black', lw=2, label='Median Predicted EDSS')
|
|
]
|
|
plt.legend(handles=legend_elements, loc='upper left', frameon=True, shadow=True, title="Legend")
|
|
|
|
# Final Labels
|
|
plt.title('Comparison of EDSS Scores Across Confidence Levels', fontsize=16, pad=20)
|
|
plt.xlabel('LLM Certainty Score (%)', fontsize=12)
|
|
plt.ylabel('EDSS Score (0-10)', fontsize=12)
|
|
plt.ylim(-0.5, 10.5)
|
|
plt.yticks(np.arange(0, 11, 1))
|
|
plt.grid(axis='y', linestyle='--', alpha=0.3)
|
|
|
|
plt.tight_layout()
|
|
plt.show()
|
|
|
|
|
|
##
|
|
|
|
|
|
|
|
# %% EDSS vs Boxplot
|
|
|
|
import pandas as pd
|
|
import numpy as np
|
|
import json
|
|
import glob
|
|
import os
|
|
import matplotlib.pyplot as plt
|
|
import seaborn as sns
|
|
from matplotlib.patches import Patch
|
|
|
|
def categorize_edss(value):
|
|
if pd.isna(value): return np.nan
|
|
elif value <= 1.0: return '0-1'
|
|
elif value <= 2.0: return '1-2'
|
|
elif value <= 3.0: return '2-3'
|
|
elif value <= 4.0: return '3-4'
|
|
elif value <= 5.0: return '4-5'
|
|
elif value <= 6.0: return '5-6'
|
|
elif value <= 7.0: return '6-7'
|
|
elif value <= 8.0: return '7-8'
|
|
elif value <= 9.0: return '8-9'
|
|
elif value <= 10.0: return '9-10'
|
|
else: return '10+'
|
|
|
|
def plot_edss_vs_confidence_boxplot(json_dir_path):
|
|
# 1. Load all Predictions
|
|
all_preds = []
|
|
json_files = glob.glob(os.path.join(json_dir_path, "*.json"))
|
|
|
|
for file_path in json_files:
|
|
with open(file_path, 'r', encoding='utf-8') as f:
|
|
data = json.load(f)
|
|
for entry in data:
|
|
if entry.get("success"):
|
|
res = entry["result"]
|
|
edss_val = pd.to_numeric(res.get('EDSS'), errors='coerce')
|
|
conf_val = pd.to_numeric(res.get('certainty_percent'), errors='coerce')
|
|
|
|
if not pd.isna(edss_val) and not pd.isna(conf_val):
|
|
all_preds.append({
|
|
'edss_cat': categorize_edss(edss_val),
|
|
'confidence': conf_val
|
|
})
|
|
|
|
df = pd.DataFrame(all_preds)
|
|
|
|
# 2. Sort categories correctly for the x-axis
|
|
cat_order = ['0-1', '1-2', '2-3', '3-4', '4-5', '5-6', '6-7', '7-8', '8-9', '9-10']
|
|
df['edss_cat'] = pd.Categorical(df['edss_cat'], categories=cat_order, ordered=True)
|
|
|
|
# 3. Plotting
|
|
plt.figure(figsize=(14, 8))
|
|
|
|
# Create Boxplot
|
|
sns.boxplot(data=df, x='edss_cat', y='confidence',
|
|
palette="Blues", width=0.6, showfliers=False)
|
|
|
|
# Add Stripplot (Dots) to show density of cases
|
|
sns.stripplot(data=df, x='edss_cat', y='confidence',
|
|
color='black', alpha=0.15, jitter=0.2, size=3)
|
|
|
|
# 4. Legend and Labels
|
|
# Since boxplot color is clear, we add a legend for the components
|
|
legend_elements = [
|
|
Patch(facecolor='#6da7d1', label='Confidence Distribution (IQR)'),
|
|
plt.Line2D([0], [0], color='black', marker='o', linestyle='',
|
|
markersize=4, alpha=0.4, label='Individual Predictions')
|
|
]
|
|
plt.legend(handles=legend_elements, loc='lower left', frameon=True)
|
|
|
|
plt.title('LLM Confidence Levels Across Clinical EDSS Categories', fontsize=16, pad=20)
|
|
plt.xlabel('Predicted EDSS Category (Clinical Severity)', fontsize=12)
|
|
plt.ylabel('Confidence Score (%)', fontsize=12)
|
|
plt.ylim(0, 105)
|
|
plt.grid(axis='y', linestyle='--', alpha=0.3)
|
|
|
|
plt.tight_layout()
|
|
plt.show()
|
|
##
|
|
|
|
|
|
|
|
|
|
# %% Correlation Boxplot
|
|
import pandas as pd
|
|
import numpy as np
|
|
import json
|
|
import glob
|
|
import os
|
|
import matplotlib.pyplot as plt
|
|
import seaborn as sns
|
|
from matplotlib.patches import Patch
|
|
from sklearn.metrics import cohen_kappa_score
|
|
|
|
def categorize_edss(value):
|
|
"""Standardized clinical categorization."""
|
|
if pd.isna(value): return np.nan
|
|
elif value <= 1.0: return '0-1'
|
|
elif value <= 2.0: return '1-2'
|
|
elif value <= 3.0: return '2-3'
|
|
elif value <= 4.0: return '3-4'
|
|
elif value <= 5.0: return '4-5'
|
|
elif value <= 6.0: return '5-6'
|
|
elif value <= 7.0: return '6-7'
|
|
elif value <= 8.0: return '7-8'
|
|
elif value <= 9.0: return '8-9'
|
|
elif value <= 10.0: return '9-10'
|
|
else: return '10+'
|
|
|
|
def plot_categorical_vs_categorical(json_dir_path, ground_truth_path):
|
|
# 1. Load Ground Truth
|
|
df_gt = pd.read_csv(ground_truth_path, sep=';')
|
|
df_gt['unique_id'] = df_gt['unique_id'].astype(str).str.strip().str.lower()
|
|
df_gt['MedDatum'] = df_gt['MedDatum'].astype(str).str.strip().str.lower()
|
|
df_gt['gt_cat'] = pd.to_numeric(df_gt['EDSS'], errors='coerce').apply(categorize_edss)
|
|
|
|
# 2. Load Predictions
|
|
all_preds = []
|
|
for file_path in glob.glob(os.path.join(json_dir_path, "*.json")):
|
|
with open(file_path, 'r', encoding='utf-8') as f:
|
|
data = json.load(f)
|
|
for entry in data:
|
|
if entry.get("success"):
|
|
res = entry["result"]
|
|
all_preds.append({
|
|
'unique_id': str(res.get('unique_id')).strip().lower(),
|
|
'MedDatum': str(res.get('MedDatum')).strip().lower(),
|
|
'pred_cat': categorize_edss(pd.to_numeric(res.get('EDSS'), errors='coerce'))
|
|
})
|
|
|
|
df_pred = pd.DataFrame(all_preds)
|
|
|
|
# 3. Merge
|
|
df_merged = pd.merge(df_pred, df_gt[['unique_id', 'MedDatum', 'gt_cat']],
|
|
on=['unique_id', 'MedDatum'], how='inner').dropna()
|
|
|
|
# 4. Set Order and Numeric Mapping for Plotting
|
|
cat_order = ['0-1', '1-2', '2-3', '3-4', '4-5', '5-6', '6-7', '7-8', '8-9', '9-10']
|
|
cat_map = {cat: i for i, cat in enumerate(cat_order)}
|
|
|
|
df_merged['gt_idx'] = df_merged['gt_cat'].map(cat_map)
|
|
df_merged['pred_idx'] = df_merged['pred_cat'].map(cat_map)
|
|
|
|
# Calculate Cohen's Kappa (Standard for categorical agreement)
|
|
kappa = cohen_kappa_score(df_merged['gt_cat'], df_merged['pred_cat'], weights='linear')
|
|
|
|
# 5. Plotting
|
|
plt.figure(figsize=(14, 8))
|
|
|
|
# BOXPLOT: Distribution of Predicted Categories relative to Ground Truth
|
|
sns.boxplot(data=df_merged, x='gt_cat', y='pred_idx',
|
|
palette="rocket", width=0.6, showfliers=False, boxprops=dict(alpha=0.5))
|
|
|
|
# STRIPPLOT: Individual counts
|
|
sns.stripplot(data=df_merged, x='gt_cat', y='pred_idx',
|
|
color='black', alpha=0.1, jitter=0.3, size=4)
|
|
|
|
# DIAGONAL REFERENCE: Perfect category match
|
|
plt.plot([0, 9], [0, 9], color='red', linestyle='--', linewidth=2)
|
|
|
|
# 6. Formatting Legend & Axes
|
|
plt.yticks(ticks=range(len(cat_order)), labels=cat_order)
|
|
|
|
legend_elements = [
|
|
Patch(facecolor='#ae3e50', alpha=0.5, label='Predicted Category Spread'),
|
|
plt.Line2D([0], [0], color='red', linestyle='--', label='Perfect Category Agreement'),
|
|
plt.Line2D([0], [0], color='black', marker='o', linestyle='', markersize=4, alpha=0.3, label='Iteration Matches'),
|
|
Patch(color='none', label=f'Linear Weighted Kappa: {kappa:.3f}')
|
|
]
|
|
plt.legend(handles=legend_elements, loc='upper left', frameon=True, shadow=True, title="Agreement Metrics")
|
|
|
|
plt.title('Categorical Agreement: Ground Truth vs. LLM Prediction', fontsize=16, pad=20)
|
|
plt.xlabel('Ground Truth Category (Clinician)', fontsize=12)
|
|
plt.ylabel('LLM Predicted Category', fontsize=12)
|
|
plt.grid(axis='both', linestyle=':', alpha=0.4)
|
|
|
|
plt.tight_layout()
|
|
plt.show()
|
|
##
|
|
|
|
|
|
|
|
# %% rainplot
|
|
import pandas as pd
|
|
import numpy as np
|
|
import json
|
|
import glob
|
|
import os
|
|
import matplotlib.pyplot as plt
|
|
import seaborn as sns
|
|
from matplotlib.patches import Patch
|
|
from matplotlib.lines import Line2D
|
|
|
|
def plot_error_distribution_by_confidence(json_dir_path, ground_truth_path):
|
|
# 1. Load Ground Truth
|
|
df_gt = pd.read_csv(ground_truth_path, sep=';')
|
|
df_gt['unique_id'] = df_gt['unique_id'].astype(str).str.strip().str.lower()
|
|
df_gt['MedDatum'] = df_gt['MedDatum'].astype(str).str.strip().str.lower()
|
|
df_gt['EDSS_gt'] = pd.to_numeric(df_gt['EDSS'], errors='coerce')
|
|
|
|
# 2. Load Predictions
|
|
all_preds = []
|
|
for file_path in glob.glob(os.path.join(json_dir_path, "*.json")):
|
|
with open(file_path, 'r', encoding='utf-8') as f:
|
|
data = json.load(f)
|
|
for entry in data:
|
|
if entry.get("success"):
|
|
res = entry["result"]
|
|
all_preds.append({
|
|
'unique_id': str(res.get('unique_id')).strip().lower(),
|
|
'MedDatum': str(res.get('MedDatum')).strip().lower(),
|
|
'EDSS_pred': pd.to_numeric(res.get('EDSS'), errors='coerce'),
|
|
'confidence': pd.to_numeric(res.get('certainty_percent'), errors='coerce')
|
|
})
|
|
|
|
df_merged = pd.merge(pd.DataFrame(all_preds), df_gt[['unique_id', 'MedDatum', 'EDSS_gt']],
|
|
on=['unique_id', 'MedDatum'], how='inner').dropna()
|
|
|
|
# 3. Calculate Error
|
|
df_merged['error'] = df_merged['EDSS_pred'] - df_merged['EDSS_gt']
|
|
|
|
# 4. Bin Confidence
|
|
bins = [0, 70, 80, 90, 100]
|
|
labels = ['Low (<70%)', 'Moderate (70-80%)', 'High (80-90%)', 'Very High (90-100%)']
|
|
df_merged['conf_bin'] = pd.cut(df_merged['confidence'], bins=bins, labels=labels)
|
|
|
|
# Calculate counts for labels
|
|
counts = df_merged['conf_bin'].value_counts().reindex(labels)
|
|
new_labels = [f"{l}\n(n={int(counts[l])})" for l in labels]
|
|
|
|
# 5. Plotting
|
|
plt.figure(figsize=(13, 8))
|
|
|
|
# Using a sequential color palette (Light blue to Dark blue)
|
|
palette_colors = sns.color_palette("Blues", n_colors=len(labels))
|
|
|
|
vplot = sns.violinplot(data=df_merged, x='conf_bin', y='error', inner="quartile",
|
|
palette=palette_colors, cut=0)
|
|
|
|
# Reference line at 0
|
|
plt.axhline(0, color='#d9534f', linestyle='--', linewidth=2.5)
|
|
|
|
# 6. UPDATED LEGEND WITH CORRECT COLORS
|
|
legend_elements = [
|
|
# Legend items for the color gradient
|
|
Patch(facecolor=palette_colors[0], label='Confidence: <70%'),
|
|
Patch(facecolor=palette_colors[1], label='Confidence: 70-80%'),
|
|
Patch(facecolor=palette_colors[2], label='Confidence: 80-90%'),
|
|
Patch(facecolor=palette_colors[3], label='Confidence: 90-100%'),
|
|
# Legend items for the symbols
|
|
Line2D([0], [0], color='black', linestyle=':', label='Quartile Lines (25th, 50th, 75th)'),
|
|
Line2D([0], [0], color='#d9534f', linestyle='--', lw=2.5, label='Zero Error (Perfect Match)')
|
|
]
|
|
|
|
plt.legend(handles=legend_elements, loc='upper left', frameon=True, shadow=True, title="Legend & Confidence Gradient")
|
|
|
|
# Formatting
|
|
plt.title('Error Magnitude vs. LLM Confidence Levels', fontsize=16, pad=20)
|
|
plt.xlabel('LLM Certainty Group', fontsize=12)
|
|
plt.ylabel('Prediction Delta (EDSS_pred - EDSS_gt)', fontsize=12)
|
|
plt.xticks(ticks=range(len(labels)), labels=new_labels)
|
|
plt.grid(axis='y', linestyle=':', alpha=0.5)
|
|
|
|
plt.tight_layout()
|
|
plt.show()
|
|
|
|
# plot_error_distribution_by_confidence('jsons_folder/', 'ground_truth.csv')
|
|
##
|
|
|
|
# %% Usage
|
|
|
|
# --- Usage ---
|
|
#plot_categorized_edss('/home/shahin/Lab/Doktorarbeit/Barcelona/Data/iteration',
|
|
# '/home/shahin/Lab/Doktorarbeit/Barcelona/Data/GT_Numbers.csv')
|
|
|
|
#plot_subcategory_analysis('/home/shahin/Lab/Doktorarbeit/Barcelona/Data/iteration', '/home/shahin/Lab/Doktorarbeit/Barcelona/Data/GT_Numbers.csv')
|
|
#plot_certainty_vs_accuracy_by_category('/home/shahin/Lab/Doktorarbeit/Barcelona/Data/iteration', '/home/shahin/Lab/Doktorarbeit/Barcelona/Data/GT_Numbers.csv')
|
|
#audit_matches('/home/shahin/Lab/Doktorarbeit/Barcelona/Data/iteration', '/home/shahin/Lab/Doktorarbeit/Barcelona/Data/GT_Numbers.csv')
|
|
|
|
|
|
#plot_edss_boxplot('/home/shahin/Lab/Doktorarbeit/Barcelona/Data/iteration', '/home/shahin/Lab/Doktorarbeit/Barcelona/Data/GT_Numbers.csv')
|
|
#plot_binned_calibration('/home/shahin/Lab/Doktorarbeit/Barcelona/Data/iteration', '/home/shahin/Lab/Doktorarbeit/Barcelona/Data/GT_Numbers.csv')
|
|
|
|
#plot_edss_vs_confidence_boxplot('/home/shahin/Lab/Doktorarbeit/Barcelona/Data/iteration')
|
|
#plot_gt_vs_llm_boxplot('/home/shahin/Lab/Doktorarbeit/Barcelona/Data/iteration', '/home/shahin/Lab/Doktorarbeit/Barcelona/Data/GT_Numbers.csv')
|
|
#plot_categorical_vs_categorical('/home/shahin/Lab/Doktorarbeit/Barcelona/Data/iteration', '/home/shahin/Lab/Doktorarbeit/Barcelona/Data/GT_Numbers.csv')
|
|
plot_error_distribution_by_confidence('/home/shahin/Lab/Doktorarbeit/Barcelona/Data/iteration', '/home/shahin/Lab/Doktorarbeit/Barcelona/Data/GT_Numbers.csv')
|
|
##
|