Discussion
In this study, we developed a strategy to incorporate machine learning
into predicting real-world time-on-treatment curves. To this end, we
generalized the problem into predicting expected future time on
treatment and then stratified the distribution of the predicted time. We
showed strong performance of this approach in predicting rwTTD across a
variety of influencing factors using simulated data. We showed its
flexibility to be applied to any machine learning base classifiers. We
then showed its robustness when trained and tested on different
populations. Lastly, we demonstrated its robust performance using real
world lung cancer and head and neck cancer data treated with
pembrolizumab.
Although rwTTD is a critical metric in monitoring the efficacy of a
treatment in the real world patient populations, no study has yet
attempted to establish machine learning models to predict rwTTD. The key
obstacle is that rather than predicting individual scores, we are
required to predict a curve. This notion and strategy is new, and will
spur the field of curve prediction in many other research fields. Of
note, we demonstrated that the aggregation of individuals does not
reflect the overall profile of the population, which is an important
rationale behind the approach we presented in this study.
This study opens the possibility of many follow-up directions. For
example, can such models be applied to clinical trial data, and using
the generated model to predict real-world populations? Can models be
well generalized from one demographic group to another? While we touched
these aspects using simulated data and real world pembrolizumab data, it
will be of interest to test in other diseases and drugs as well. How
does the interpolation function affect the performance of the model? How
do other base learners such as deep learning, Gaussian Progress
Regression work with this model? Our approach allows incorporation of
any supervised base learner which can be tested in future studies
concerning other diseases and therapeutics. Finally, this study opens
the possibility of population-wise predictions, which is distinguished
from individual-wise prediction. This will have enormous applications in
the future in all research areas whose current focus is on individual
predictions.