In scalar value, we predicted diagnoses and symptoms

In this paper, we address the task of learning predictive models for pediatric patients’ hospital readmission problem. Hospital readmission is a critical metric of quality and cost of healthcare which is regarded as very challenging and important for both medical and machine learning community. In contrast to the prediction of readmission, where the label is a single scalar value, we predicted diagnoses and symptoms with which patient will readmit.

We tried to answer two questions. Namely, does information about the structure of labels can improve the predictive performance of models, where the structure is both expert-driven and data-driven? If so, can we derive some conclusion when expert-driven or data-driven structure over outputs will provide better performance? In order to answer these questions, we used Predictive Clustering Trees which provides decision tree like model which are suited for structured output multi-label prediction. Based on results we can conclude that inclusion of structure over output space, both expert-driven and data-driven, does improve the performance of Predictive Clustering Trees. However, performances with expert-driven hierarchy have better performance for label-based and ranking-based evaluations measures, while PCT with data-driven hierarchies showed better performance in example-based measures.

We Will Write a Custom Essay Specifically
For You For Only $13.90/page!


order now

Limitation of this study is the usage of ICD-9 codes is made for reimbursement purposes and do not necessarily describes the clinical condition of the patient accurately. However, there is no consensus about the formalization of medical concepts. Use of expert-driven ontologies can be very limited since they aim to provide formalized knowledge about medical concepts that should apply to the overall population. However, development of such formalized knowledge repositories is a very challenging task because of the complexity of medical phenomena, duration, and cost of Randomized Controlled Trials etc. This often leads to inadequate treatments for specific patients, and an increase of costs cite{van2016randomized}. On the other side, data-driven hierarchies (ontologies) can provide close insights into a specific population that is being analyzed. This way data-driven ontologies can confirm or disprove hypothesis provided by domain knowledge for specific patients or groups. Again, there is a number of challenges that prevent development of highly accurate data driven models that could be exploited in medical practice to their full potential (heterogeneity of data, class and feature imbalance, sparsity, lack of data, privacy concerns)

In future work, we plan to include more expert-driven hierarchies in order to check domain knowledge and derive conclusions which hierarchy is better suited for the problem at hand. Further, we will develop a methodology for fusion of data and knowledge drove hierarchy that will be based on quantification of compliance. Finally, we plan to derive a methodology for PCT which utilizes hierarchical structure over input space also. With this improvement we expect to gain new features which have better generalization and therefore partitioning of space will generate rules with higher coverage leading to more general models and very possibly better performance.