Towards Accurate and Interpretive Disease Diagnosis for Aging Adults Through Deep Learning Framework

Population aging has become a serious problem all around the world. The world's population is aging faster than ever before. What’s worse is that as the population ages, a serious problem arises: as the number of elderly people increases, medical resources begin to become scarce. For example, during the COVID-19 pandemic, many elderly people died due to lack of treatment and other medical resources such as masks. Similar to the COVID-19, some diseases are believed to be closely related to age and lead to higher mortality rates among the elderly. It is becoming more important to establish an efficient and accurate diagnostic system for such diseases. In this way, the impact of population aging on society can also be further alleviated. The rapid development of artificial intelligence technology can bring a turning point to this problem. Deep learning has become one of the most current topics in the field of artificial intelligence. With the swelling data of the Internet, deep learning showed emerging power to solve problems from various subjects that traditional machine learning methods could not handle well. Disease diagnosis is exactly one of them. In this research, We are aimed to build an accurate and interpretive disease diagnosis model leveraging the deep learning framework, specifically the Transformer, to diagnose the physical conditions of the elderly and reveal the relationship between specific microbiomes and disease. The key to the problem lies in what the classification model uses as features. Human gut microbiome has been proved to involve in a variety of chemical and biological processes in the human body and have complex connection with our health including specific diseases. In fact, sampling gut microbiome from patients to do disease diagnosis has been expected to be a more efficient and non-invasive method comparing to existed method such as blood test. Furthermore, the rapid development of gene sequencing technology also provides strong hardware support for this idea such as High-throughput sequencing. The core idea of this study is to use the relative abundance of human intestinal microorganisms as features to train a high-precision classification model for specific disease diagnosis. We evaluated five algorithms on six different and independent datasets with three age-related diseases. The results showed that Transformer could show the best performance in most of the cases, although it demands more computing resources to train. Furthermore, we managed to improve the generalization ability with the MK-MMD method. Finally, we did biomarker discovery to all the models with decent performance by SHAP and student t-test.