Tsinghua University develops advanced AI doctor training system
A team of researchers at Tsinghua University has recently unveiled a fully simulated environment for AI medical training. The virtual environment enables virtual doctors to train in a virtual hospital without the need for real-life interaction with actual patients.
Also Read: AI helps farmers plant crops and raise livestock
The Intelligent Industry Research Institute (AIR) and the Department of Computer Science and Technology at Tsinghua University created a highly detailed simulation of a hospital that includes virtual workers and patients. The virtual environment, known as the Agent Hospital, allows AI doctors to carry out diagnosis and treatment on thousands of virtual patients. The AI doctors, owing to the process of learning and making mistakes, gradually became proficient in diagnosing and treating patients.
Virtual environment enables AI doctors to diagnose virtual patients
The main benefit of this simulated environment is that it allows the user to work with a large number of cases without having to wait for real patients. This method is not only faster in terms of training but also economically efficient. In this way, the AI can accumulate thousands of virtual patients’ data in a relatively short time.
Also Read: Meta suspend plans to train AI on EU users’ data
The researchers employed a process known as the MedAgent-Zero method in the training of virtual AI doctors on 10,000 patients’ records. They trained large language models on eight diseases to produce electronic health records. These diseases included acute nasopharyngitis, acute rhinitis, bronchial asthma, chronic bronchitis, COVID-19, Influenza A, Influenza B, and mycoplasma infection. The virtual patients had different signs and stages of the disease, making the database of the training set diverse.
The AI doctor, built with the GPT-3.5-turbo-1106 model, became much more effective in a short time. It handled 10,000 virtual cases and reported high success rates in patient examination, diagnosis, and treatment, indicating that the system had a good learning curve. The overall success rates of the specific disease ranged from 88% for examination to 95.6% for diagnosis and 77. 6% for treatment.
GPT-4 outperforms GPT-3 on medical licensing exam questions
In the subsequent study, the Tsinghua researchers applied the MedAgent-Zero method to the gpt-4-1106-preview model. The performance comparison was done on 1273 questions from the MedQA dataset that replicates medical licensing tests like USMLE. The study revealed a marked increase with the GPT-4 model with a success rate of 93. 06% for respiratory disease questions as opposed to 84.72% for GPT-3.
Also Read: New AI Training Program for 1,000 African Cybersecurity Entrepreneurs
The successful development and implementation of the Agent Hospital and MedAgent-Zero method by Tsinghua University marks a new revolution in medical training. Virtual simulations are beneficial as they allow AI doctors to practice in a controlled and adjustable setting and thus, make the training process more efficient.
Cryptopolitan Reporting by Brenda Kanana
