Diagnosis Recommendation Using Machine Learning Scientific Workflows

摘要

Diagnosis recommendation plays a significant role in healthcare, where a clinician infers an optimal diagnosis for a patient. This problem has a major impact on improving patients' quality of life. Existing machine learning techniques for solving this problem require many labeled instances, which are not readily available. To overcome this limitation, in this paper, we present a scientific workflow for representing a semisupervised clustering based diagnosis recommendation model. In this approach, initial clusters are formed from a labeled dataset; then imposing certain relative threshold to a cluster, frequent patterns and their corresponding labels are obtained. Subsequently, unlabeled instances are labeled by assigning them to the most similar clusters. Finally, we form clusters on the generated new datasets and recommend the diagnosis label by applying a certain minimum threshold. To evaluate our model, we perform extensive experiments on the i2b2 datasets and compared our proposed algorithms with the self-training and co-training methods. The experimental results show that our proposed algorithm outperforms the mentioned methods in most cases. The proposed workflow is implemented in the DATAVIEW system.