Multimodality-aware Intra-dialytic Hypotension Classifier: A Bilingual NLP Approach to Classify Dialysis Records

Nursing records, rich in free-text data, encapsulate valuable insights into patients' physiological status and physicians' diagnoses, offering more holistic perspectives than mere tabular data. The scrutiny of these free-text descriptions often enables physicians to rapidly assimilate the patients' historical and current health states, thereby facilitating accurate diagnosis and treatment. One particularly pressing issue is the identification of intra-dialytic hypotension (IDH) symptoms within these records. IDH, a sporadic yet potentially fatal complication during hemodialysis sessions for end-stage kidney disease patients, is diagnosed by a decrease of 20 mmHg in systolic blood pressure during hemodialysis or a drop of 10 mmHg in mean arterial pressure coupled with hypo-perfusion-related symptoms. In response to this, we have meticulously developed a Multi-modality-aware IDH Classifier (MMIDHC) using a natural language processing (NLP) model aimed at the classification of these records. Primarily, this model focuses on the classification of English-Chinese bilingual dialysis records concerning IDH. Composed of three distinct modules, the model's intricate processes are delineated in Figure 1.

Free-text Pre-processing: To investigate the influence of long sentences and noise on classification, we implemented sentence segmentation. Additionally, we enhanced the contextual relationships and augmented the sample size by merging adjacent segments in both antegrade and retrograde fashions, as context ordered-shuffling.

Free-text Feature Extraction: To expedite the process and avoid the need to rebuild NLP models, we adopted publicly available NLP models that support the Chinese language for feature extraction. In this context, we utilized the Multilingual Universal Sentence Encoder (mUSE) proposed by Yang et al. [3] and the Multilingual Sentence BERT (mSBERT) introduced by Reimers et al. [2]. The latter model builds upon Sentence BERT (SBERT) [1] and incorporates the Teacher-Student Model. Both these multilingual NLP models serve as encoders, parsing free-text, and converting natural language into feature vectors. The performance of these models, in terms of metric indicators, showed minimal differences at a fixed threshold. Nevertheless, the mSBERT encoder consistently outperformed mUSE in speed, approximately 20 times faster. Given the potential impact of encoder speed on the operation of the multimodule model, we selected mSBERT as the primary encoder.

Classification Learning: We employed a Multilayer Perceptron (MLP) for classifying the extracted features. In order to improve the reliability of our classification, we incorporated the MLP's Intradialytic Hypotension (IDH) results with the blood pressure readings from the ongoing dialysis session into an ensemble model. This method facilitates more nuanced decision-making. For example, if the MLP detects IDH symptoms (Yes), but the blood pressure reading is normal (False), we classify the result as False owing to lack of fulfilling definition of IDH. However, if both the MLP and the blood pressure readings suggest a normal state, we categorize the dialysis record as Non-IDH. Our baseline model has achieved encouraging results across all metrics: Accuracy 0.951, Recall 0.955, Precision 0.913, and F1-score 0.932. For more information, please visit our website: https://github.com/IlikeBB/Far-Eastern-Memorial-Hospital-IDH_RecommendationSystem_Project.