Анализ медицинских врачебных заключений¶

!cd /user/Demo_public/biomedical/analis_medical_result

 !pip install -r /user/Demo_public/biomedical/analis_medical_result/requirements.txt

import pandas as pd
import seaborn as sns
import matplotlib.pyplot as plt
from sklearn.metrics import cohen_kappa_score
from statsmodels.stats.inter_rater import fleiss_kappa
import numpy as np
from sklearn.metrics import accuracy_score, recall_score, confusion_matrix, balanced_accuracy_score, mean_squared_error
from sklearn.metrics import precision_recall_curve, average_precision_score, f1_score, fbeta_score, roc_auc_score,ConfusionMatrixDisplay, roc_curve, auc
from sklearn.preprocessing import MinMaxScaler
from scipy.stats import sem, t

focus = pd.read_csv('/user/Demo_public/biomedical/analis_medical_result/Дата.csv')

focus.head()

focus = focus.drop('ID Файла', axis=1)

focus_corr = focus.corr(method='spearman')
filtered = focus_corr.where(focus_corr > 0.5).abs()

focus_corr_ = focus.corr(method='spearman')
filtered_ = focus_corr_.where(focus_corr < 0.1)

focus_corr_mean = focus_corr.mean(axis=1).sort_values(ascending=False)
focus_corr_mean

plt.figure(figsize=(8,8))
sns.heatmap(focus_corr.round(2), annot=True, center=0, cmap='inferno')
plt.title('Корреляция между разметками врачей')

plt.figure(figsize=(8,8))
sns.heatmap(filtered.round(2), annot=True, center=0, cmap='inferno')
plt.title('Корреляция между разметками врачей с средней по силе корреляции')

confidence = focus.sum(axis=1).sort_values(ascending=False) / focus.shape[1] * 100
confidence.head(20)

average_marking = focus.mean(axis=1)

average_doctor_deviation = focus.sub(average_marking, axis=0).abs().mean().sort_values(ascending=False)
average_doctor_deviation

covariance_matrix = focus.cov()
covariance_matrix

covariance_matrix_mean = covariance_matrix.mean(axis=1).sort_values(ascending=False)
covariance_matrix_mean

doctor_13 = focus['Врач№13']
doctor_3 = focus['Врач№3']
doctor_12 = focus['Врач№12']
doctor_10 = focus['Врач№10']

cohen_kappa_score(doctor_13, doctor_3), cohen_kappa_score(doctor_12, doctor_10)

counts = np.zeros((focus.shape[0], 2), dtype=int)
for idx, row in focus.iterrows():
  counts[idx, 0] = (row == 0).sum()
  counts[idx, 1] = (row == 1).sum()

fleiss_kappa(counts)

kappa_scores = pd.DataFrame(focus.columns, focus.columns)

for col in focus.columns:
  for col_ in focus.columns:
    kappa_scores.loc[col, col_] = cohen_kappa_score(focus[col], focus[col_])
kappa_scores = kappa_scores.drop([0], axis=1)

kappa_scores_mean = kappa_scores.mean(1).sort_values(ascending=False)
kappa_scores_mean

standard_deviation_mean=focus.sub(focus.mean(axis=1), axis=0).pow(2).mean().sort_values(ascending=False)
standard_deviation_mean

std_mean=focus.std().sort_values(ascending=False)
std_mean

diagnosis_major = pd.Series([0 if x > y else 1 for x, y in zip(counts[:, 0], counts[:, 1])])

accuracy_df = focus.apply(lambda x: accuracy_score(diagnosis_major, x)).sort_values(ascending=False)
accuracy_df

recall_df = focus.apply(lambda x: recall_score(diagnosis_major, x)).sort_values(ascending=False)
recall_df

columns = focus.columns
confusion_matrices = {}
for column in columns:
  confusion_matrix_ = confusion_matrix(diagnosis_major, focus[column])
  confusion_matrices[column] = confusion_matrix_.sum() - np.diag(confusion_matrix_).sum()

errors_df = pd.DataFrame(list(confusion_matrices.items()), columns=['Врач', 'Количество ошибок']).sort_values(ascending=False, by='Количество ошибок')
errors_df

counts[:, 0].sum(), counts[:, 1].sum()

balanced_df = focus.apply(lambda x: balanced_accuracy_score(diagnosis_major, x)).sort_values(ascending=False)
balanced_df

colors = ['blue', 'green', 'red', 'purple', 'orange', 'yellow', 'brown', 'pink', 'gray', 'cyan', 'magenta', 'lime', 'teal', 'navy']

fig, axes = plt.subplots(nrows=5, ncols=2,figsize=(13, 24))
axes = axes.flatten()
axes[0].bar(focus_corr_mean.index.tolist(), focus_corr_mean.values.tolist(), color=colors[0])
axes[0].set_title('Средняя корреляция по врачам')
axes[0].set_xlabel('Врач')
axes[0].set_ylabel('Средняя корреляция')
axes[0].tick_params(axis='x', rotation=45)
axes[1].bar(average_doctor_deviation.index.tolist(), average_doctor_deviation.values.tolist(), color=colors[1])
axes[1].set_title('Среднее отклонение по врачам')
axes[1].set_xlabel('Врач')
axes[1].set_ylabel('Среднее отклонение')
axes[1].tick_params(axis='x', rotation=45)
axes[2].bar(covariance_matrix_mean.index.tolist(), covariance_matrix_mean.values.tolist(), color=colors[2])
axes[2].set_title('Средняя ковариация по врачам')
axes[2].set_xlabel('Врач')
axes[2].set_ylabel('Средняя ковариация')
axes[2].tick_params(axis='x', rotation=45)
axes[3].bar(kappa_scores_mean.index.tolist(), kappa_scores_mean.values.tolist(), color=colors[3])
axes[3].set_title('Средняя оценка по Капо-Коэно по врачам')
axes[3].set_xlabel('Врач')
axes[3].set_ylabel('Средняя оценка по Капо-Коэно')
axes[3].tick_params(axis='x', rotation=45)
axes[4].bar(standard_deviation_mean.index.tolist(), standard_deviation_mean.values.tolist(), color=colors[4])
axes[4].set_title('Среднеквадратическое отклонение по врачам')
axes[4].set_xlabel('Врач')
axes[4].set_ylabel('Среднеквадратическое отклонение')
axes[4].tick_params(axis='x', rotation=45)
axes[5].bar(std_mean.index.tolist(), std_mean.values.tolist(), color=colors[5])
axes[5].set_title('Среднее СКО по врачам')
axes[5].set_xlabel('Врач')
axes[5].set_ylabel('Среднее СКО')
axes[5].tick_params(axis='x', rotation=45)
axes[6].bar(accuracy_df.index.tolist(), accuracy_df.values.tolist(), color=colors[6])
axes[6].set_title('Мажоритарная точность по врачам')
axes[6].set_xlabel('Врач')
axes[6].set_ylabel('Мажоритарная точность')
axes[6].tick_params(axis='x', rotation=45)
axes[7].bar(recall_df.index.tolist(), recall_df.values.tolist(), color=colors[7])
axes[7].set_title('Мажоритарная полнота по врачам')
axes[7].set_xlabel('Врач')
axes[7].set_ylabel('Мажоритарная полнота')
axes[7].tick_params(axis='x', rotation=45)
axes[8].bar(errors_df['Врач'], errors_df['Количество ошибок'], color=colors[8])
axes[8].set_title('Ошибки по врачам')
axes[8].set_xlabel('Врач')
axes[8].set_ylabel('ошибки по полнота')
axes[8].tick_params(axis='x', rotation=45)
axes[9].bar(balanced_df.index.tolist(), balanced_df.values.tolist(), color=colors[9])
axes[9].set_title('Мажоритарная сбалансированная точность по врачам')
axes[9].set_xlabel('Врач')
axes[9].set_ylabel('Мажоритарная сбалансированная точность')
axes[9].tick_params(axis='x', rotation=45)
plt.subplots_adjust(hspace=0.6)
plt.tight_layout()

fig, axes = plt.subplots(figsize=(7, 7))
axes.bar(accuracy_df.index.tolist(), accuracy_df.values.tolist(), color=colors[6], alpha=1, label='Мажоритарная точность')
axes.set_title('Мажоритарная точность и сбалансированная точность по врачам')
axes.set_xlabel('Врач')
axes.set_ylabel('Мажоритарная точность, сбалансированная точность')
axes.tick_params(axis='x', rotation=45)
axes.bar(balanced_df.index.tolist(), balanced_df.values.tolist(), color=colors[9], alpha=0.5, label='Мажоритарная сбалансированная точность')
axes.legend()
axes.set_ylim(0, 1.25)
plt.tight_layout()

plt.figure(figsize=(8, 6))
sns.boxplot(data=confidence, orient='v', color='lightblue')
plt.ylabel('Вероятность наличия патологии (%)')
plt.title('Распределение вероятностей наличия патологии')
plt.tight_layout()
plt.show()

metrics_df = pd.concat([focus_corr_mean, average_doctor_deviation, covariance_matrix_mean, kappa_scores_mean, standard_deviation_mean, std_mean, accuracy_df, recall_df, balanced_df, errors_df.set_index('Врач')['Количество ошибок']], axis=1)
metrics_df.columns=['focus_corr_mean', 'average_doctor_deviation', 'covariance_matrix_mean', 'kappa_scores_mean', 'standard_deviation_mean', 'std_mean', 'accuracy_df', 'recall_df', 'balanced_df', 'errors_df']

metrics_df

scaler = MinMaxScaler()
normalized_df = pd.DataFrame(scaler.fit_transform(metrics_df), index=metrics_df.index, columns=metrics_df.columns)

normalized_df

weights = {
    'kappa_scores_mean': 0.3,
    'covariance_matrix_mean': -0.2,
    'average_doctor_deviation': -0.25,
    'focus_corr_mean': 0.25,
    'balanced_accuracy': 0.25,
    'recall': 0.25,
    'accuracy': 0.25,
    'std_mean': -0.15,
    'standard_deviation_mean': -0.10,
    'errors_df': -0.2
}

normalized_df['final_score'] = (normalized_df * pd.Series(weights)).sum(axis=1)

normalized_df['final_score'].sort_values(ascending=False)

plt.figure(figsize=(10, 6))
sns.barplot(x=normalized_df['final_score'].sort_values(ascending=False).values, y=normalized_df['final_score'].sort_values(ascending=False).index, palette='coolwarm')
plt.title('Рейтинг врачей')
plt.xlabel('Рейтинг')
plt.ylabel('Врач')
plt.show()

votes = focus.sum(axis=1)
f = focus * normalized_df['final_score']
final_votes = (f.sum(1) > 0.4).astype(int)
confidence__ = focus.sum(axis=1)
weights_ = {
    'doctor' : 0.4,
    'confidence': 0.6
}

Diagnosis_weights = ((confidence__ / focus.shape[1]) * weights_['confidence'] + (f.sum(1) > 0.4).astype(int) * weights_['doctor'])
Diagnosis = (Diagnosis_weights >= 0.36).astype(int)

Diagnosis.sum()

normalized_diagnosis_weights = (Diagnosis_weights - np.min(Diagnosis_weights)) / (np.max(Diagnosis_weights) - np.min(Diagnosis_weights))
confidence = normalized_diagnosis_weights * 100
confidence.sort_values(ascending=False)

mean = np.mean(Diagnosis_weights)
std_err = sem(Diagnosis_weights)

confidence_level = 0.95
n = len(Diagnosis_weights)
h = std_err * t.ppf((1 + confidence_level) / 2, n - 1)

lower_bound = mean - h
upper_bound = mean + h

# Вывод доверительного интервала
print(f"Доверительный интервал: [{lower_bound}, {upper_bound}]")

df = pd.concat([Diagnosis, confidence], axis=1)
df.columns = ['Diagnosis', 'Probability, %']
df['Probability, %'] = df.apply(lambda row: 100 - row['Probability, %'] if row['Diagnosis'] == 0 else row['Probability, %'], axis=1)
df

mean_probability = np.mean(Diagnosis_weights)
print(f"Средняя вероятность: {mean_probability}")

# Стандартное отклонение вероятностей
std_deviation = np.std(Diagnosis_weights)
print(f"Стандартное отклонение вероятностей: {std_deviation}")

# Максимальная и минимальная вероятность
max_probability = np.max(Diagnosis_weights)
min_probability = np.min(Diagnosis_weights)
print(f"Максимальная вероятность: {max_probability}")
print(f"Минимальная вероятность: {min_probability}")

accuracy_d = accuracy_score(diagnosis_major, Diagnosis)
recall_d = recall_score(diagnosis_major, Diagnosis)
balanced_accuracy_d = balanced_accuracy_score(diagnosis_major, Diagnosis)
F1_d = f1_score(diagnosis_major, Diagnosis)
fbeta_d = fbeta_score(diagnosis_major, Diagnosis, beta=0.5)
roc_auc_d = roc_auc_score(diagnosis_major, Diagnosis)
print(f"Точность: {accuracy_d}")
print(f"Полнота: {recall_d}")
print(f"Сбалансированная Точность: {balanced_accuracy_d}")
print(f"F1: {F1_d}")
print(f"Fbeta: {fbeta_d}")
print(f"roc-auc: {roc_auc_d}")

confusion_matrix_ = confusion_matrix(diagnosis_major, Diagnosis)
disp = ConfusionMatrixDisplay(confusion_matrix_)
disp.plot(cmap=plt.cm.Blues)
plt.title('Confusion Matrix')

precision, recall, _ = precision_recall_curve(diagnosis_major, Diagnosis)
average_precision = average_precision_score(diagnosis_major, Diagnosis)

plt.figure()
plt.step(recall, precision, where='post', color='b', alpha=0.2, linestyle='-', linewidth=2, label='Precision-Recall curve')
plt.fill_between(recall, precision, step='post', alpha=0.2, color='b')
plt.xlabel('Recall')
plt.ylabel('Precision')
plt.title('Precision-Recall curve: AP={0:0.2f}'.format(average_precision))
plt.legend(loc="lower left")