Patient Outcome and Zero-shot Diagnosis Prediction with Hypernetwork-guided Multitask Learning

Proceedings of the 17th Conference of the European Chapter of the Association for Computational Linguistics, pages 589–598 May 2-6, 2023 ©2023 Association for Computational Linguistics Patient Outcome and Zero-shot Diagnosis Prediction with Hypernetwork-guided Multitask Learning Shaoxiong Ji and Pekka Marttinen Aalto University, Finland {shaoxiong.ji; pekka.marttinen}@aalto.fi Abstract Multitask deep learning has been applied to patient outcome prediction from text, taking clinical notes as input and training deep neu- ral networks with a joint loss function of multiple tasks. However, the joint training scheme of multitask learning suffers from inter-task interference, and diagnosis predic- tion among the multiple tasks has the gener- alizability issue due to rare diseases or unseen diagnoses. To solve these challenges, we pro- pose a hypernetwork-based approach that gen- erates task-conditioned parameters and coeffi- cients of multitask prediction heads to learn task-specific prediction and balance the mul- titask learning. We also incorporate semantic task information to improve the generalizabil- ity of our task-conditioned multitask model. Experiments on early and discharge notes ex- tracted from the real-world MIMIC database show our method can achieve better perfor- mance on multitask patient outcome predic- tion than strong baselines in most cases. Be- sides, our method can effectively handle the scenario with limited information and improve zero-shot prediction on unseen diagnosis cate- gories. 1 Introduction Recent advances apply artificial intelligence to pre- dict clinical events or infer the probable diagno- sis for clinical decision support (Shickel et al., 2017; Li et al., 2021). Those works extensively study clinical data from Electronic Health Records (EHRs). For example, the Doctor AI model for pre- dictive modeling on EHRs (Choi et al., 2016) and the Deep Patient model for unsupervised patient record representation learning (Miotto et al., 2016) is proposed by leveraging longitudinal medical data. Within the patient records, we can also find various clinical notes written by general practitioners or physicians and annotated with specific diagnosis re- sults, medical codes, and other clues of patient out- come. Those clinical notes consist of meaningful health information, including health profile, clini- cal synopsis, diagnostic investigations, and medi- cations, which can empower the clinical decision making (Li et al., 2021). Clinical text mining has demonstrated its feasibility in diverse healthcare applications, including medical code prediction (Ji et al., 2021), readmission prediction (Huang et al., 2019) and diagnosis prediction (van Aken et al., 2021). Readmission and length-of-stay prediction can help with utility management, especially when healthcare service is in high demand. Diagnostic prediction can support clinicians in making deci- sions based on many clinical reports. This study focuses on multitask patient outcome prediction from clinical notes, which classifies patients’ di- agnostic results and predicts clinical outcomes to empower expert clinicians and improve healthcare services’ efficiency. Timeline Health Profile Adverse Reactions and Alerts Clinical Synopsis Operative Report Blood Test Medications Radiology Report Xray Examination Findings Findings Results MedicationDirections Duration Results Findings Summary and Report Multitask Deep Learning Model Figure 1: An example of multitask patient outcome pre- diction based on sequential inputs of clinical notes in the electronic health …

Content truncated to 500 words