Multi-disease Diagnosis System Based On Hierarchical Classification And Deep Reinforcement Learning

In the medical dialogue and consultation system,the dialogue strategy and disease diagnosis module play a decisive role in completing the consultation task.The dialogue strategy affects whether the diagnosis model can obtain high-value symptoms to give accurate diagnosis results,and the reasoning results of the diagnosis model are also important information to help the dialogue strategy judge whether to ask and the tendency to ask.However,there are the following difficulties in actual work.First,the multi-disease diagnosis effect of the traditional flat classification scheme will decrease with the increase of the number of diseases.Second,the symptoms of various diseases in the real data overlap,resulting in differences between diseases.The similarities and differences of diseases are not the same,so the difficulty of diagnosis of each disease is very different.Third,the system should locate the key symptoms and ensure the accuracy of diagnosis in the shortest possible dialogue rounds.First of all,the concept of hybrid disease separable degree and its calculation method are proposed to realize the measurement of the degree of separability between two diseases,which provide a basis for subsequent disease stratification.Secondly,combined with hybrid disease separable degree,a multi-level diagnostic framework is constructed based on agglomerative hierarchical clustering algorithms,and the original large-scale disease diagnosis task is resolved into multiple small-scale disease cluster classification tasks.Thirdly,the criterion of minimizing dynamic conditional entropy is proposed,and the symptoms that can minimize the disease conditional entropy are selected,thereby reducing the uncertainty of diagnostic results in as few rounds as possible.Finally,based on reinforcement learning algorithm,a multi-disease diagnostic system is obtained by combining the multi-level diagnostic framework and the criterion of minimizing dynamic conditional entropy.This paper organizes a dataset based on the actual consultation records of the hospital.Experimental results show that hybrid disease separable degree achieves accurate measurement of disease separability,the multi-level diagnostic framework can improve the diagnostic accuracy of large-scale diseases,and can improve the classification performance of the basic model.According to the symptoms located by the criterion of minimizing the dynamic conditional entropy,the rounds of symptom inquiry can be shortened,but it needs to be matched with a suitable disease diagnosis model.The multi-disease diagnosis system is more accurate in locating symptoms,and the diagnostic accuracy is not inferior to most models.