Research: Development of AI-powered 'digital twins' improves accuracy of medical predictions

The system's accuracy was tested in three areas: predicting the condition of 16,500 patients with non-small cell lung cancer, 35,000 patients in intensive care units, and 1,100 people with Alzheimer's disease. In all cases, the new model outperformed existing solutions, with an average prediction error of 1.3–3.4%. Importantly, DT-GPT not only more accurately predicted individual indicators but also preserved the relationships between them and the overall data structure. This makes the model a more reliable tool for mapping patient dynamics and advances the practical use of DT in medicine.

Furthermore, DT-GPT is capable of operating on a zero-learn basis. This means the system can predict parameters not included in the initial configuration. In the study, it outperformed traditional algorithms on 13 of 69 such parameters. Another feature is its robustness to typical problems in electronic medical records: the model maintains high performance even with a large number of omissions, random errors, and typos in medical records.

However, the technology has limitations: it copes less well with rare and abrupt changes in state, and can also inherit distortions in the original data. Therefore, widespread implementation will require the involvement of specialists and refinement of methods that improve accuracy and reliability.

Nevertheless, as the researchers concluded, the use of digital data can accelerate clinical trials, help doctors make decisions, improve the quality of personalized medicine, and open up new opportunities for the digital health market – from disease modeling services to predictive analytics and patient monitoring tools.

The perception of digital data is also actively debated in society. A study by the University of Zurich found that 62% of respondents are interested in using such technologies, but 87% oppose their mandatory implementation. Meanwhile, 79% believe that only medical professionals should have access to the data, and 64% are willing to share anonymized information freely for scientific purposes.