VIENNA — A machine learning model using routinely collected laboratory parameters estimates low-density lipoprotein cholesterol (LDL-C) with reasonable accuracy, potentially offering a practical and low-cost approach to cardiovascular risk screening.

Developed using nearly 300,000 patient records, the model could help identify individuals with elevated LDL who may be missed in screening programs that rely on total cholesterol alone.

“Using simple, routinely measured parameters, we can predict the level of low-density lipoprotein,” said Valentin Kokorin, MD, Dr Sci, professor of internal medicine at the People’s Friendship University of Russia, Moscow. “It needs no cost at all, no additional resources. You can just use the calculator and have the result.”

The findings were presented at 24th European Congress of Internal Medicine (ECIM) 2026.

Machine Learning Bridges Screening Gap

LDL-C is a major driver of cardiovascular risk, but it is not always measured in routine screening. “In our system, we usually evaluate only total cholesterol,” Kokorin explained. “LDL is an additional test, which is more complicated and more expensive.”

That limitation may leave some patients undetected. “Some patients have normal total cholesterol, but at the same time low-density lipoprotein is already elevated,” he said. “So we can have a new strategy of searching for high-risk patients.”

The model was trained on anonymized data from 279,332 individuals, incorporating commonly available variables including blood counts, total cholesterol, fasting glucose, age, and sex. “We made quite a big machine learning program with these data and trained the machine and then used it for external validation,” Kokorin explained.

Performance was strong, with 92% accuracy in the test dataset and 87% in external validation. The mean absolute percentage error was between 7% and 9%. “There is some error in prediction, about less than 10%,” he said. “But we believe this is acceptable.”

He stressed that the model is not intended to replace laboratory testing but to guide it. “If you see that the values are high or outside the range, then you can send the patient to the selected test,” he said.

Toward Earlier, Lower-Cost Prevention

The work reflects a broader shift toward earlier identification of risk. “The concept of medicine now is not to treat advanced stages of disease but to try to find them at an early stage and prevent severe conditions,” Kokorin noted.

Machine learning, he suggested, offers a way to do that using data already collected in routine care. “This is a huge opportunity to search for high-risk patients who will be outside of simple screening,” he said, adding that the tool is already available as an online calculator, currently in Russian, with plans for wider access.

“This is not only about LDL,” he added. “We are working on prediction of ferritin levels and some other tests not only inside cardiology.”

‘A Good Tool but We Must Control It’

Kokorin emphasized that machine learning should support, not replace, clinical judgment. “Yes, it is a good tool, a new opportunity for doctors,” he said. “But of course, we should control the machine because otherwise we will have problems.”

He added that the approach aligns with broader pressures on healthcare systems. “In every healthcare system, ministries are searching for something cheaper and easier to save resources,” he said. “This is a solution in this case.”

Taken together, the findings suggest that machine learning could make cardiovascular screening more efficient, particularly where resources are constrained. “We can just use existing data and find patients earlier,” Kokorin said.

Useful, but Depends on Context

Commenting on the findings, Pietro Di Francesco, MD, University of Milan, Milan, Italy, said the approach is attractive, particularly in terms of cost. “When we can derive data without having to actually take the test, it is always a good thing to do because we could reduce cost and otherwise we have to test all the population,” he said.

However, he added that, in many clinical settings, LDL is already measured alongside total cholesterol. “In our practice when we ask for total cholesterol, we often also ask for LDL,” he said. “So I wonder whether the algorithm could predict LDL without even having total cholesterol.”

Even so, he said the model could be useful where testing is more limited. “I think it is still interesting because LDL testing takes more reagents and more resources.”

Kokorin and Di Francesco reported having no relevant financial relationships.