Clinical Outcome Prediction from Admission Notes using Self-Supervised Knowledge Integration

Proceedings of the 16th Conference of the European Chapter of the Association for Computational Linguistics, pages 881–893 April 19 - 23, 2021. ©2021 Association for Computational Linguistics 881 Clinical Outcome Prediction from Admission Notes using Self-Supervised Knowledge Integration Betty van Aken1, Jens-Michalis Papaioannou1, Manuel Mayrdorfer2, Klemens Budde2, Felix A. Gers1 and Alexander L¨oser1 1Beuth University of Applied Sciences Berlin 2Charit´e Berlin {bvanaken,michalis.papaioannou,gers,aloeser}@beuth-hochschule.de {manuel.mayrdorfer,klemens.budde}@charite.de Abstract Outcome prediction from clinical text can prevent doctors from overlooking possible risks and help hospitals to plan capacities. We simulate patients at admission time, when de- cision support can be especially valuable, and contribute a novel admission to discharge task with four common outcome prediction targets: Diagnoses at discharge, procedures performed, in-hospital mortality and length-of-stay predic- tion. The ideal system should infer outcomes based on symptoms, pre-conditions and risk factors of a patient. We evaluate the effec- tiveness of language models to handle this scenario and propose clinical outcome pre- training to integrate knowledge about patient outcomes from multiple public sources. We further present a simple method to incorpo- rate ICD code hierarchy into the models. We show that our approach improves performance on the outcome tasks against several baselines. A detailed analysis reveals further strengths of the model, including transferability, but also weaknesses such as handling of vital values and inconsistencies in the underlying data. 1 Introduction Clinical professionals make decisions about pa- tients under strong time constraints. The patient information at hand is often unstructured, e.g. in the form of clinical notes written by other medical personnel in limited time. Clinical decision support (CDS) systems can help in these scenarios by point- ing towards related cases or certain risks. Clinical outcome prediction is a fundamental task of CDS systems, in which the patient’s development is pre- dicted based on data from their Electronic Health Record (EHR). In this work we focus on textual EHR data available at admission time. Figure 1 shows a sample admission note with high- lighted parts that – according to medical doctors – must be considered when evaluating a patient. Encoding clinical notes with pre-trained language models. Neural models need to extract relevant facts from such notes and learn complex relations between them in order to associate certain clinical outcomes. Pre-trained language models such as BERT (Devlin et al., 2019) have shown to be able to both extract information from noisy text and to capture task-specific relations in an end-to-end fashion (Tenney et al., 2019; van Aken et al., 2019). We thus base our work on these models and pose the following questions: • Can pre-trained language models learn to pre- dict patient outcomes from their admission information only? • How can we integrate knowledge about out- comes that doctors gain from medical litera- ture and previous patients? • How well would these models work in clinical practice? Are they able to interpret common risk factors? Where are they failing? Simulating patients at admission time. Exist- ing work on text-based outcome prediction focuses on progress notes after a certain time …

Content truncated to 500 words