C-index with censoring adjustment (IPCW)

[BBolosSierra]

Hi,

Thanks for your work!

I was evaluating the discrimination performance, particularly the concordance index with your package.
Setting censor_surv = "km" or censor_surv = None with EvalSurv in the particular case of concordance index, does not seem to change the results. Which makes sense theoretically with the definition of concordance in your implementation. Could you confirm that IPCW (adjusting for censoring distribution) with EvalSurv is not intended for the concordance index implemented?

Thank you so much in advance,

I have tested it like the following:

# checking if there is issues with ipcw
import numpy as np
import pandas as pd
import h5py
from lifelines import CoxPHFitter
from lifelines.utils import concordance_index
import requests

from pycox.evaluation import EvalSurv

# Load mb
url = "https://github.com/jaredleekatzman/DeepSurv/raw/refs/heads/master/experiments/data/metabric/metabric_IHC4_clinical_train_test.h5"
file_name = "metabric_IHC4_clinical_train_test.h5"

with open(file_name, 'wb') as f:
    f.write(requests.get(url).content)

# Read train and test
with h5py.File(file_name, 'r') as f:
    x_train = f['train']['x'][()]
    e_train = f['train']['e'][()]
    t_train = f['train']['t'][()]
    x_test = f['test']['x'][()]
    e_test = f['test']['e'][()]
    t_test = f['test']['t'][()]


print(x_train.shape)
print(e_train.shape) 
print(t_train.shape) 

df_train = pd.DataFrame(x_train)
df_train['time'] = t_train + 1e-8
df_train['status'] = e_train

df_test = pd.DataFrame(x_test)
df_test['time'] = t_test + 1e-8 
df_test['status'] = e_test

# Fit cox
cox = CoxPHFitter()
cox.fit(df_train, duration_col='time', event_col='status')

baseline_surv = cox.baseline_survival_
times_grid = baseline_surv.index.values

# Get risk
risk_scores = cox.predict_partial_hazard(df_test)

# Predict on the test sample
surv_probs = np.array([baseline_surv.values.flatten() ** rs for rs in risk_scores]).T
surv_df = pd.DataFrame(surv_probs, index=times_grid)

# Conver to right type
times_test = df_test['time'].values
events_test = df_test['status'].values.astype(int)

# Check with and without the censoring
ev_no_ipcw = EvalSurv(surv_df, times_test, events_test, censor_surv=None)
ev_ipcw = EvalSurv(surv_df, times_test, events_test, censor_surv="km")

print("C-index without IPCW:", ev_no_ipcw.concordance_td())
print("C-index with IPCW (KM):", ev_ipcw.concordance_td())

Thanks again,
Best,