1: Binary classification model using data that exclude all patients lost to follow-up [12, 13] (e.g., exclude any patient not observed for the full time-at-risk)
|
The labels are correct as we observed all the patients in the training data for the complete time-at-risk follow-up
|
We reduce the size of the training data (the longer the time-at-risk, the smaller the dataset)
|
If the health outcome is often fatal, then we may exclude all or the majority of the patients who have the health outcome
|
May limit model generalizability to only those who are healthy
|
2: Binary classification model using data that include all patients (including those lost to follow-up) [14] (e.g., include every patient in the cohort. A patient not observed for the full time-at-risk is included but their outcome is determined based on whether they experienced the outcome during the observed time-at-risk)
|
We do not compromise generalizability
|
Labels may be incorrect for those who are lost to follow-up (this noise may impact the model’s ability to learn)
|
Larger sample size
|
3: Binary classification model using data that exclude patients lost to follow-up unless they have the outcome prior to loss to follow-up [15] (e.g., only exclude patients not observed for the full time-at-risk if they did not have the outcome during the observed time-at-risk. This means patients with a partial time-at-risk who have the outcome during this time are still included)
|
The labels are correct
|
Generalizability may be compromised
|
We include all outcomes
|
Outcome patients may be sicker as we can include those who die within time-at-risk but this is not possible for non-outcomes
|
Do not lose outcomes when outcome is associated to death
|
4: Cox model using data that includes all patients (including those lost to follow-up) [16] (e.g., include every patient, even those not observed for the full time-at-risk. The survival time is the minimum of time to end of observation, time to outcome or time-at-risk end (time to study period end from cohort index)
|
Method suitable for censored patients
|
Not intended for risk prediction, the main purpose is hazard rate calculation per predictor. Requires baseline hazard function for prediction
|
Predict survival time (time before event) rather than risk of event
|
Computationally more expensive
|