以異質關係進行疾病診斷碼表示法學習之研究

Representation Learning for Diagnosis Codes by Heterogeneous Relationships

羅仕翰(Lo Shih-Han)

指導教授 ： 柯佳伶

國立臺灣師範大學/理學院/資訊工程學系/碩士(2020年)

https://doi.org/10.6345/NTNU202000433

摘要

疾病診斷碼表示法的學習在近年來被廣為研究，然而許多研究僅考慮疾病診斷碼間的出現關聯資料。本論文以疾病診斷碼為主體，同時考慮與其他診斷碼、個人屬性特徵、醫療用藥或醫療處置等資料出現的關聯進行表示法的學習，用於下一次看診的疾病診斷碼進行預測。本論文提出兩種訓練表示法的方法，第一種是對各屬性特徵分開進行獨立訓練，第二種是將特徵合在一個模型中進行聯合訓練。表示法訓練完成後，針對兩種不同的表示法訓練方法所得到的疾病診斷碼表示法提供對應的預測模型，其中針對獨立訓練的疾病診斷碼表示法提出三種整合方式：直接接合、權重合成及注意力機制。實驗結果顯示，獨立訓練模型的直接接合及注意力機制整合方式，以及聯合訓練模型，與Med2Vec相較起來，在預測效能都有顯著的上升。特徵組合探討方面，以聯合訓練模型，特徵採用疾病診斷碼搭配看診時間及醫療處置時，可得到最佳預測效果。

關鍵字

疾病診斷碼預測 ； 獨立訓練 ； 聯合訓練 ； 整合模型

並列摘要

Representation learning of diagnosis codes is studied by extensive research, but most of them only use co-occurrence data among disease codes. This thesis not only takes the disease codes, but also uses the occurrence data with other diseases, the personal features, and the medical or procedure treatments for representation learning, which are used to predict the diagnosis codes occurring in the next visit. We propose two methods for representation learning of diagnosis codes. The first one is an independent model to train the representation of diagnosis codes by each feature separately. The second one uses all features to jointly train their representation in the same model, which is called the joint learning approach. Moreover, for the learnt representations of codes, there are different prediction models designed. Among them, three integrational methods are proposed to combine the representations learnt from independent model in the prediction model, i.e., concatenation, combination with weights, and attention mechanism. The results of experiments show that the performances of independent model with concatenation, independent model with attention mechanism and joint learning model are better than Med2Vec significantly. In terms of feature combination, the best predictive effect is obtained by the joint training model by using disease diagnosis codes, diagnosis time, and medical procedures.

並列關鍵字

Diagnosis Codes Prediction ； Independent Training Model ； Joint Learning ； Integration Model

參考文獻

[1] D. Bahdanau, K. Cho, and Y. Bengio. “Neural Machine Translation by Jointly Learning to Align and Translate.” Paper presented at Proceedings of the 3rd International Conference on Learning Representations (ICLR), 2015.

[2] E. Choi, M. T. Bahadori, E. Searles, C. Coffey, M. Thompson, J. Bost, and J. Tejedor-Sojo, and J. Sun. “Multi-Layer Representation Learning for Medical Concepts.” Paper presented at Proceedings of the 22nd ACM SIGKDD International Conference on Knowledge Discovery and Data Mining (KDD), 2016.

[3] E. Choi, M. T. Bahadori, L. Song, W. F. Stewart, and J. Sun. “GRAM: Graph-based Attention Model for Healthcare Representation Learning.” Paper presented at Proceedings of the 23rd ACM SIGKDD International Conference on Knowledge Discovery and Data Mining (KDD), 2017.

[4] E. Choi, M. T. Bahadori, and J. Sun. “Doctor AI: Predicting Clinical Events via Recurrent Neural Networks.” In 2016 Machine Learning and Healthcare Conference (MLHC), 2016.

[5] E. Choi, M. T. Bahadori, J. Sun, J. Kulas, A. Schuetz, and W. F. Stewart. “RETAIN: An Interpretable Predictive Model for Healthcare using Reverse Time Attention Mechanism.” Paper presented at Advances in Neural Information Processing Systems (NIPS), 2016.