RECOMED: A comprehensive pharmaceutical recommendation system

To build datasets containing useful information from drug databases and recommend a list of drugs to physicians and patients with high accuracy by considering a wide range of features of people, diseases, and chemicals.

A comprehensive pharmaceutical recommendation system was designed based on the features of people, diseases, and medicines extracted from two major drug databases and the created datasets of patients and drug information. Then, the recommendation was given based on recommender system algorithms using patient and caregiver ratings and the knowledge obtained from drug specifications and interactions. Sentiment analysis was employed by natural language processing approaches in pre-processing, along with neural network-based methods and recommender system algorithms for modelling the system. Patient conditions and medicine features were used to make two models based on matrix factorization. Then, we used drug interaction criteria to filter drugs with severe or mild interactions with other drugs. We developed a deep learning model for recommending drugs using data from 2304 patients as a training set and 660 patients as our validation set. We used knowledge from drug information and combined the model's outcome into a knowledge-based system with the rules obtained from constraints on taking medicine.

Our recommendation system can recommend an acceptable combination of medicines similar to the existing prescriptions available in real life. Compared with conventional matrix factorization, our proposed model improves the accuracy, sensitivity, and hit rate by 26 %, 34 %, and 40 %, respectively. In addition, it improves the accuracy, sensitivity, and hit rate by an average of 31 %, 29 %, and 28 % compared to other machine learning methods. We have open-sourced our implementation in Python.

Compared to conventional machine learning approaches, we obtained average accuracy, sensitivity, and hit rates of 31 %, 29 %, and 28 %, respectively. Compared to conventional matrix factorisation our proposed method improved the accuracy, sensitivity, and hit rate by 26 %, 34 %, and 40 %, respectively. However, it is acknowledged that this is not the same as clinical accuracy or sensitivity, and more accurate results can be obtained by gathering larger datasets.