USING ARTIFICIAL INTELLIGENCE TO IMPROVE HEALTHCARE QUALITY AND EFFICIENCY

Abstract

In recent years, artificial intelligence (AI), especially machine learning (ML) and deep learning (DL), has represented one of the most exciting advances in science. The performance of ML-based AI in many areas, such as computer vision, voice recognition, and natural language processing has improved dramatically, offering unprecedented opportunities for application in a variety of different domains. In the critical domain of healthcare, great potential exists for a broader application of ML to improve quality and efficiency. At the same time, there are substantial challenges in the development and implementation of AI in healthcare.

This dissertation aims to study the application of state-of-the-art AI technologies in healthcare, ranging from original method development to model interpretation and real-world implementation. First, a novel DL-based method is developed to efficiently analyze the rich and complex electronic health record data. This DL-based approach shows promise in facilitating the analysis of real-world data and can complement clinical knowledge by revealing deeper insights. Both knowledge discovery and performance of predictive models are demonstrably boosted by this method.

Second, a recurrent neural network (named LSTM-DL) is developed and shown to outperform all existing methods in addressing an important real-world question, patient cost prediction. A series of novel analyses is used to derive a deeper understanding of deep learning’s advantages. The LSTM-DL model consistently outperforms other models with nearly the same level of advantages across different subgroups. Interestingly, the advantage of the LSTM-DL is significantly driven by the amount of fluctuation in the sequential data. By opening the “black box,” the parameters learned during the training period are examined, and is it demonstrated that LSTM-DL’s ability to react to high fluctuation is gained during the training rather than inherited from its special architecture. LSTM-DL can also learn to be less sensitive to fluctuations if the fluctuation is not playing an important role.

Finally, the implementation of ML models in real practice is studied. Since at its current stage of development, ML-based AI will most likely assistant human workers rather than replace them, it is critical to understand how human workers collaborate with AI. An AI tool was developed in collaboration with a medical coding company, and successfully implemented in the real work environment. The impact of this tool on worker performance is examined. Findings show that use of AI can significantly boost the work productivity of human coders. The heterogeneity of AI’s effects is further investigated, and results show that the human circadian rhythm and coder seniority are both significant factors in conditioning productivity gains. One interesting finding regarding heterogeneity is that the AI has its best effects when a coder is at her/his peak of performance (as opposed to other times), which supports the theory of human-AI complementarity. However, this theory does not necessarily hold true across different coders. While it could be assumed that senior coders would benefit more from the AI, junior coders’ productivity is found to improve more. A further qualitative study uncovers the underlying mechanism driving this interesting effect: senior coders express strong resistance to AI, and their low trust in AI significantly hinders them from realizing the AI’s value.