Real-World Medical Knowledge Graph and Clinical Event Graph Construction

Based on the disease probability knowledge graph, the application of intelligent diagnosis will be more convenient. We apply it in clinical decision support scenarios. When a patient first visits the hospital, the doctor writes the chief complaint, current medical history, and other information in the case. Using NLP technology, symptoms and characteristics of onset from the patient’s chief complaint and current medical history are extracted and input into an improved Bayesian model to infer the most likely list of diseases for the patient. Once the doctor selects a diagnosis, the system automatically recommends examination and testing items based on the doctor’s selected diagnosis, with the recommendations sourced from the constructed medical knowledge graph.

During hospitalization, patients use many medications; for example, a diabetic patient may use dozens of medications during hospitalization. When doctors want to check the patient’s main medications during hospitalization or after discharge, the traditional approach is to flip through medication orders from front to back. Based on the above knowledge graph, we match the medications related to the patient’s main diagnosis in the knowledge graph with the medications used during hospitalization, using the PSR indicator to rank the medications. As shown, the main medications for this type 2 diabetes patient are Repaglinide tablets, Voglibose tablets, and Aspart insulin injection. This greatly saves the doctor’s time and improves work efficiency.

In the field of medical informatics, predictions of the next medication for patients are often based on the public database MIMIC. Since it is time-series data from the ICU, deep learning algorithms based on Bi-LSTM as the basic model have achieved good results. We added a graph embedding layer to the Bi-LSTM model to predict the next medication for patients, and experimental results show that both the convergence speed and accuracy of the model have improved. This indicates that the combination of graph embedding and deep learning is very helpful for model optimization. However, prediction models based on vector representation are often challenged due to their poor interpretability in serious medical scenarios such as drug recommendations, making them more suitable for medical text NLP and intelligent triage scenarios.

Further Reading