The 'medical artificial intelligence (AI)' used for disease diagnosis, early disease prediction, and medical image interpretation is rapidly penetrating the medical field as an assistant to doctors. Recently, AI for early disease prediction has gained attention for its potential to advance treatment timing, increase patient survival rates, and reduce medical expenses.

Amidst this, European scientists have developed an AI capable of predicting the likelihood of over 1,000 diseases. This innovative achievement represents a significant advancement in the medical AI field, surpassing the utility of existing disease prediction AIs, which were limited to predicting one or two types of diseases.

● Prediction of Over 1,000 Diseases

A joint research team, including the German Cancer Research Center, European Bioinformatics Institute, and Swiss Federal Institute of Technology Zurich, developed an AI model called 'Delphi-2M' that can predict an individual's lifelong health changes by analyzing vast medical record data. This was published in the international journal 'Nature' on the 17th (local time).

The research team developed Delphi-2M based on the AI model 'GPT-2' developed by OpenAI. Initially, the model was trained with 80% of the medical data from 400,000 individuals in the UK Biobank from 2006 to 2020. The UK Biobank is a large-scale dataset that has collected long-term data on past disease diagnoses, disease occurrences, health status, genomics, and lifestyle of approximately 500,000 UK citizens. Delphi-2M used 2.2 million numbers (weights) for its training.

The research team verified the performance of Delphi-2M using the remaining 20% of the UK Biobank data and the 'Danish National Disease Registry' medical data of 1.93 million Danish citizens. They input personal medical data up to the age of 60 into Delphi-2M and simulated when each of over 1,000 diseases would occur over the next 20 years, creating virtual future disease prediction data.

The comparison between the virtual data and actual data of individuals aged 70 to 75 showed similarities. The diagnostic accuracy (AUC) averaged 0.76, and the disease prediction accuracy 10 years later reached 0.7. AUC closer to 1 indicates higher accuracy.

● Explanation of AI Judgments... Increased Transparency

Another reason this research result is noteworthy is that it not only predicts over 1,000 diseases but also analyzes the judgment process and basis of Delphi-2M. This is an attempt to secure trust and transparency in AI judgments, which is a controversial issue in recent medical AI. AI judgments are directly linked to human lives. Currently, AI can make judgments but cannot explain the basis for them.

The research team analyzed the reasons for Delphi-2M's judgments using the 'Shapley Additive Explanations (SHAP)' method. SHAP is a technique that fairly evaluates how much each variable, viewed as a player, contributes to the predicted value. According to the analysis, the basis for Delphi-2M's judgment of high pancreatic cancer risk was the conclusion that a history of digestive diseases increases the risk of pancreatic cancer by 19 times. Additionally, it was observed that mental illnesses such as depression and dementia, and childbirth and pregnancy-related diseases often appear together in groups. Delphi-2M captured patterns consistent with previously known medical results, making judgments based on them. It accurately identified significant relationships between diseases, and between health status or lifestyle and specific diseases.

The research team stated, "Delphi-2M helps identify individuals at high risk of diseases and supports long-term planning for personalized medical services such as screening plans," adding, "The research results do not stop at 'high diagnostic prediction accuracy' but also show which factors increased the risk, securing transparency."

The transparency of medical AI was also emphasized in a review paper published on the 10th in 'Nature Reviews Bioengineering' by Professor Lee Soo-in of the Department of Computer Science and Engineering at the University of Washington, who gained attention for developing the SHAP methodology. Professor Lee, who was recognized with the Samsung Ho-Am Prize last year, is researching technologies that can explain AI judgments in fields where accountability is important, such as healthcare, finance, and law.

In the paper, Professor Lee analyzed various papers and cases dealing with the transparency of medical AI and pointed out that for medical AI to be introduced in actual practice, the data used for model construction, the learning process, and the performance test results must be transparently disclosed. In addition to transparency, as medical AI continues to advance, discussions are ongoing about complementing AI judgments with medical staff intervention, ensuring data diversity, continuous monitoring, and regulation.

Popular News