Patient-Specific Prediction with Bayesian Personalized Decision PathsJohnson, Adriana LJ (2021) Patient-Specific Prediction with Bayesian Personalized Decision Paths. Doctoral Dissertation, University of Pittsburgh. (Unpublished) This is the latest version of this item.
AbstractMachine learning algorithms can be useful in predicting patient outcomes under uncertainty. Many algorithms employ “population” methods to optimize a single, static model to predict well on average for an entire population, but such models may perform poorly for patients who differ greatly from the average patient or majority of the population. Personalized methods seek to optimize predictive performance for every patient by tailoring a patient-specific model to each individual. Prior work on personalized methods includes clustering methods like k-nearest neighbor and tree-derived methods like decision paths. It has been shown in multiple domains that ensembles of decision trees often outperform single decision tree models by reducing variance, capturing a range of significant features, and mitigating the uncertainty involved in model selection. However, ensemble methods have been used only sparingly in the context of personalized models. The use of Bayesian scoring in model construction has also been shown to improve predictive performance of decision tree and decision path algorithms. In this dissertation, we developed and evaluated several novel personalized decision path methods – including methods that construct single personalized decision paths as well as methods that construct ensembles of paths that use Bayesian scoring in the form of personalized random forest and personalized boosted trees. We found that the use of a random forest ensemble approach was associated with improvements to the predictive performance of personalized decision paths in terms of discrimination and calibration, and the use of Bayesian scoring was associated with improvements to the predictive performance of personalized decision paths, decision trees, and random forest ensembles of decision trees. However, we did not observe a global performance benefit from using personalization, ensemble approaches, and Bayesian scoring together compared to corresponding population and non-Bayesian algorithmic methods. Share
Details
Available Versions of this Item
MetricsMonthly Views for the past 3 yearsPlum AnalyticsActions (login required)
|
![[feed]](/images/feed-icon-32x32.png){kind=icon}