import torch.nn.functional as F
cv_test_index = y_test.index
cv_predicted_df_dict = {}
folds = [0]
for ep in np.arange(29,30,1):
print(ep)
for fold in folds:
cv_test_predicted_df = pd.DataFrame(np.zeros((len(cv_test_index),2)), index=cv_test_index, columns=['score','gold'])
file = np.sort(os.listdir(f"{folder}/{str(fold)}"))[ep] # Get the last model file in the folder
print(f"Fold {fold} model:", file)
model_path = f"{folder}/{str(fold)}/{file}"
checkpoint = torch.load(model_path, map_location=torch.device('cpu'))
model_ft = models.resnet101().to(device)
model_test = nn.Sequential(model_ft, net_add).to(device)
model_test.load_state_dict(checkpoint['model_state_dict'])
model_test.cuda()
criterion = nn.CrossEntropyLoss()
epoch_test = checkpoint['epoch']
loss_test = checkpoint['loss']
model_test.eval()
loss_epoch_test=[]
y_proba = []
y_gold = []
y_pred = []
with torch.no_grad():
for b, (X, y) in enumerate(test_dataloader):
outputs = model_test(X.cuda())
_, preds = torch.max(outputs, 1)
y_proba += torch.flatten(outputs[:,1]).cpu().tolist()
y_gold += y.data.cpu().tolist()
y_pred += preds.cpu().tolist()
# loss = criterion(outputs.float(), torch.tensor([[1,0] if label==0 else [0,1] for label in y_test]).to(device).float())
# loss_epoch_test.append(loss.item())
auc=roc_auc_score(y_gold,y_proba)
print(f"Fold {fold} AUROC: {auc}")
cv_test_predicted_df.loc[cv_test_index,'score'] = y_proba
cv_test_predicted_df.loc[cv_test_index,'gold'] = y_gold
cv_predicted_df_dict[fold] = cv_test_predicted_df
# Average the scores of the 5 folds
scores_sum=0
for f in folds:
scores_sum += cv_predicted_df_dict[f]["score"]
cv_test_predicted_df["score"] = scores_sum/len(folds)
# cv_test_predicted_df["score"] = cv_predicted_df_dict[0]["score"]
cv_test_predicted_df_ihc = cv_test_predicted_df.join(meta_erihc, how='inner')
cv_test_predicted_df_ihc["Subtype"] = None
cv_test_predicted_df_ihc["Subtype"][(cv_test_predicted_df_ihc["er_status_by_ihc"]=='Positive') & (cv_test_predicted_df_ihc["her2_status_by_ihc"]=='Negative')] = "ER+/HER2-"
# cv_predicted_df_ihc["Subtype"][cv_predicted_df_ihc["her2_status_by_ihc"]=='Positive'] = "ERBB2+"
cv_test_predicted_df_ihc["Subtype"][(cv_test_predicted_df_ihc["er_status_by_ihc"]=='Negative') & (cv_test_predicted_df_ihc["her2_status_by_ihc"]=='Negative') & (cv_test_predicted_df_ihc["pr_status_by_ihc"]=='Negative')] = "TNBC"
cv_test_predicted_df_ihc["EERES"] = brca_earlyes_es.loc[cv_test_predicted_df_ihc.index]
cv_test_predicted_df_ihc["ESR1"] = brca_lv3.loc[cv_test_predicted_df_ihc.index]["ESR1"]
brca_clinical_testing = brca_clinical.join(cv_test_predicted_df_ihc, how='inner')
brca_clinical_testing_dfs = brca_clinical_testing[["DFS_MONTHS", "DFS_STATUS", "score", "Subtype", "EERES", "ESR1"]].dropna()
brca_clinical_testing_dss = brca_clinical_testing[["DSS_MONTHS", "DSS_STATUS", "score", "Subtype", "EERES", "ESR1"]].dropna()
brca_clinical_testing_pfs = brca_clinical_testing[["PFS_MONTHS", "PFS_STATUS", "score", "Subtype", "EERES", "ESR1"]].dropna()
brca_clinical_testing_os = brca_clinical_testing[["OS_MONTHS", "OS_STATUS", "score", "Subtype", "EERES", "ESR1"]].dropna()
plt.scatter(cv_test_predicted_df_ihc["EERES"], cv_test_predicted_df_ihc["score"])
plt.xlabel("EERES")
plt.ylabel("Score")
r, p = pearsonr(cv_test_predicted_df_ihc["EERES"], cv_test_predicted_df_ihc["score"])
plt.title(f"Spearman R: {r}, p-value: {p}")
auroc = roc_auc_score(cv_test_predicted_df_ihc["gold"], cv_test_predicted_df_ihc["score"])
RocCurveDisplay.from_predictions(
cv_test_predicted_df_ihc["gold"],
cv_test_predicted_df_ihc["score"],
name=f"EERES>{eeres_threshold:.3f}",
color="darkorange",
)
# plt.plot([0, 1], [0, 1], "k--", label="chance level (AUC = 0.5)")
plt.axis("square")
plt.xlabel("False Positive Rate")
plt.ylabel("True Positive Rate")
# plt.title("Receiver Operating Characteristic")
plt.legend()
plt.show()
brca_clinical_testing_erposerbb2neg = brca_clinical.join(cv_test_predicted_df_ihc[cv_test_predicted_df_ihc['Subtype']=="ER+/HER2-"], how='inner')
brca_clinical_testing_erposerbb2neg
brca_clinical_testing_erposerbb2neg_dfs = brca_clinical_testing_erposerbb2neg[["DFS_MONTHS", "DFS_STATUS", "score", "EERES", "ESR1"]].dropna()
brca_clinical_testing_erposerbb2neg_dss = brca_clinical_testing_erposerbb2neg[["DSS_MONTHS", "DSS_STATUS", "score", "EERES", "ESR1"]].dropna()
brca_clinical_testing_erposerbb2neg_pfs = brca_clinical_testing_erposerbb2neg[["PFS_MONTHS", "PFS_STATUS", "score", "EERES", "ESR1"]].dropna()
brca_clinical_testing_erposerbb2neg_os = brca_clinical_testing_erposerbb2neg[["OS_MONTHS", "OS_STATUS", "score", 'EERES', 'ESR1']].dropna()
finding_best_threshold_with_FS_SS(brca_clinical_testing_erposerbb2neg_dfs, brca_clinical_testing_erposerbb2neg_dss, ["DFS", "DSS"], 'score', brca_clinical_testing_erposerbb2neg_dss['score'].quantile(0.1), brca_clinical_testing_erposerbb2neg_dss['score'].quantile(0.9), "Predicted Score")
finding_best_threshold_with_FS_SS(brca_clinical_testing_erposerbb2neg_pfs, brca_clinical_testing_erposerbb2neg_os, ["PFS", "OS"], 'score', brca_clinical_testing_erposerbb2neg_os['score'].quantile(0.1), brca_clinical_testing_erposerbb2neg_os['score'].quantile(0.9), "Predicted Score")
brca_clinical_score = brca_clinical.join(cv_test_predicted_df_ihc, how='inner')[['PFS_MONTHS','PFS_STATUS','ESR1','score','age_at_diagnosis','ajcc_pathologic_tumor_stage']].dropna()
cph_pfs = CoxPHFitter()
cph_pfs.fit(brca_clinical_score[['PFS_MONTHS','PFS_STATUS','score']], duration_col='PFS_MONTHS', event_col='PFS_STATUS')
cph_pfs.print_summary() # access the individual results using cph.summary
