The researchers of the three organizations reveal the GluFormer, a creative artificial intelligence model (Generative AI) based on a transformer architecture similar toGPT From home OpenAI – but instead of text – it produces a forecast of blood sugar levels and other health indicators, based on monitoring past data
Diabetics - or others who are required to monitor their sugar level - often deal with the question of how eating a cookie or pastry will affect them. Now, a new creative artificial intelligence model is able to predict the answer: researchers from the Weizmann Institute of Science, the startup company Pheno.AI and the NVIDIA research group in Israel, led the development of GluFormer - a foundation model that is able to predict the future sugar levels of patients , as well as other health indicators, relying on the patient's sugar data from the past.
Information collected from continuous glucose monitoring (CGM) systems can help diagnose patients with pre-diabetes or diabetes more quickly, according to Harvard Health Publishing and-NYU Langone Health. GluFormer's AI capabilities are able to enhance this diagnostic capability, helping therapists and patients identify anomalies, predict clinical trial outcomes, and predict medical and health metrics up to four years in advance.
Accurate predictions of future sugar levels for patients in risk groups for developing diabetes may allow doctors and patients to adopt preventive treatment strategies earlier, to predict response to drugs and treatments in clinical trials, to predict medical indicators in patients about four years in advance, and to reduce the economic effects of dealing with diabetes, which may reach to-About 2.5 trillion dollar In 2030.
In addition, the researchers showed how, after feeding the model with information related to the patient's diet, it is possible to predict how his sugar levels will respond to specific foods and changes in the diet - which can enable personalized nutrition with a high degree of accuracy. Beyond sugar levels, GluFormer can predict medical values including adipose tissue - a measure of the amount of fat in the body around organs such as the liver and pancreas; Systolic blood pressure, which is related to the risk of diabetes, and the apnea index - a measurement of sleep apnea, which is related to type 2 diabetes.
"Medical data, and in particular continuous monitoring of sugar levels, can be seen as sequences of diagnostic tests that follow biological processes throughout life," said Prof. Gal Chechik, senior director at NVIDIA and director of NVIDIA's artificial intelligence research center in Israel. "We found that the "Transformer" artificial intelligence architecture, which was developed for long text sequences, can take a sequence of medical tests and predict the results of the next test. In doing so, the model learns about how the diagnostic indicators develop over time."
"Two important factors came together at the same time to make the research we did possible: the maturation of generative artificial intelligence technologies powered by NVIDIA, alongside large-scale health data collection by the Weizmann Institute," said lead study author Guy Lutzker, an artificial intelligence researcher at -NVIDIA and PhD student at the Weizmann Institute. "This puts us in a unique position to produce and extract important medical insights from the medical indices."
AI tools like GluFormer have potential future to help hundreds of millions of people dealing with diabetes. Today, diabetes affects about 10% of the world's population, and researchers estimate that until the year 2050 Its influence will be multiplied and it will affect the lives of over 1.3 billion people worldwide. This is a disease that is among ten The world's leading causes of death, and it causes significant complications among many patients such as kidney damage, vision damage and heart problems.
The technology behind Gluformer works in the same architecture as large language models such as OpenAI's GPT - only in this case the product is sugar levels, as opposed to text in GPT. The model was trained on monitoring the blood sugar levels of over 10,000 non-diabetic study participants over 14 days, with data collected every 15 minutes using a wearable monitoring device, and was able to predict patterns of blood sugar levels. In total, the model was trained on 10 million sugar level data.
The research and development of GluFormer, carried out on Nvidia graphics processors (GPUs), was led by Prof. Eran Segal, from the Computer Science Department at the Weizmann Institute of Science, Prof. Gal Chechik, director of the Nvidia Research Center in Israel, and from the Computer Science Department at Bar University Ilan, Hagai Rossman, manager of the research team at Pheno.AI, Gal Sapir, researcher at the data science team at Pheno.AI, and the lead researcher, Guy Lutzker, a researcher at NVIDIA and a doctoral student at the Weizmann Institute.
More of the topic in Hayadan: