Comparing Ordered Logistic Regression and Random Forest for Predicting Lacosamide Trough Concentration Tiers in a Real-World Cohort: A Methodological Cautionary Study

In the world of medicine, some drugs work like a precise key fitting into a lock, while others behave more like water poured into a bucket with a hole that changes size for every person. Lacosamide is one of those drugs that varies wildly from patient to patient. It is a medication used to control seizures, and doctors rely on a process called therapeutic drug monitoring to make sure the amount of the drug in a patient's blood is just right. If the level is too low, the seizures might return; if it is too high, the patient could suffer from side effects. The challenge is that the same dose does not produce the same blood level in different people. Factors like age, body weight, and whether a patient is taking other medications can all shift the balance. For years, doctors have hoped that by feeding these known facts into a computer, they could predict exactly where a patient's drug level would fall, allowing for perfect, personalized dosing from the very start.

A team of researchers at Zhejiang University School of Medicine decided to test this hope with a large, real-world experiment. They gathered data from nearly a thousand patients, ranging from toddlers to elderly adults, who had their lacosamide levels checked at a single hospital. The researchers wanted to see if they could use two different types of mathematical tools to sort these patients into three groups: those with low drug levels, those with levels in the safe, therapeutic range, and those with levels that were too high. One tool was a traditional statistical method known as ordered logistic regression, which looks for straight-line relationships between factors like age and drug levels. The other was a machine learning technique called a random forest, a more complex system that tries to find hidden patterns and connections that simple math might miss. The goal was to see if either tool could accurately guess a patient's drug tier based on a single blood test and a list of routine details like their weight and other medications.

The researchers fed the computer systems a specific set of information for each patient: their age, sex, body weight, the dose of lacosamide they were taking adjusted for their weight, whether they were taking other seizure medications that are known to speed up the body's processing of drugs, and how many other seizure drugs they were on at the same time. They defined the "other medications" very carefully, looking only at prescriptions filled within a week of the blood test to ensure the timing matched. They then split their data, using most of it to teach the models and the rest to test how well the models performed on patients they had never seen before. The results were clear and somewhat disappointing for those hoping for a quick computational fix. Neither the traditional statistical model nor the advanced machine learning system could predict the drug levels with any useful accuracy. The models performed only slightly better than random guessing. While the machine learning model did edge out the traditional one by a small margin, both failed to reach the level of reliability needed to guide clinical decisions.

Despite the failure to predict individual levels, the study did uncover one very specific and reliable signal. The researchers found that when a patient was taking certain other seizure medications that speed up drug metabolism, their lacosamide levels were significantly lower. This effect was strong enough to be detected even when the models struggled with everything else. The data showed that patients taking these interacting drugs had about half the odds of having a high drug level compared to those who were not. This confirmed a known biological fact: these specific drugs act as a chemical accelerator, clearing lacosamide from the body faster. However, this single factor was not enough to make the overall prediction work. The models could not account for the vast amount of other variation that exists between people.

The study suggests that the problem is not the type of math used, but the nature of the data itself. A single blood test, no matter how many details are attached to it, does not contain enough information to predict where a patient's drug level will land. The researchers noted that factors they could not measure, such as whether a patient actually took their medication as prescribed, or genetic differences in how their liver processes drugs, likely play a huge role. The machine learning model did not find any secret patterns that the simpler model missed; both agreed that age and the weight-adjusted dose were the most important factors, yet even these were insufficient. The study also confirmed that the relationship between age and drug levels is not the same for everyone, with children showing a tighter link between the dose they receive and the level they achieve compared to adults, who have much more unpredictable results.

Ultimately, the research serves as a cautionary tale for the field of personalized medicine. It demonstrates that while we have powerful tools to analyze data, we cannot yet predict complex biological outcomes from a single snapshot in time. The idea that a computer algorithm could instantly tell a doctor the perfect dose based on a routine blood test and a few basic facts is not supported by this evidence. Instead, the findings reinforce the need for doctors to rely on careful observation over time, watching how a patient's levels change with repeated tests. The only clear rule that emerged is that if a patient is taking specific interacting medications, their levels will likely be lower, and this must be accounted for manually. For the rest, the path to truly individualized dosing requires more than just better software; it requires a deeper look at the patient's history, their genetics, and their actual adherence to the treatment plan, gathered over a long period rather than a single moment.