How artificial intelligence may save on tests for detecting life-threatening antibodies
In a study conducted at Ichilov Hospital, the possibility of using data from two routine coagulation tests was examined to screen early for samples likely to be negative, and to reduce unnecessary tests without forgoing a full investigation in suspicious cases. Dr. Bentzi Katz, who co-led the study, explains how artificial intelligence can find new value in results already obtained in the laboratory and assist in smarter and more efficient use of existing information.
Seven coagulation tests, on average, to finally reach one answer: negative. This is what happened in the vast majority of the 7,454 investigations we examined for the detection of lupus anticoagulant. In fact, almost 90% of them ended without the detection of the finding we were looking for. This number caused us to stop and ask: is it really necessary to perform the entire series of tests almost every time? And what if it is possible to use the information already obtained at the beginning of the investigation to identify in advance samples that are very likely to turn out negative?
Before reaching an answer, one must understand what is being looked for at all. Lupus anticoagulant is an important part of the laboratory investigation of antiphospholipid syndrome (APS). The investigation is intended to identify the characteristic effect of autoantibodies on the coagulation system. Antiphospholipid syndrome is an autoimmune disease in which the immune system produces autoantibodies against phospholipids, which are important components of cell membranes in our body. These antibodies can disrupt the normal activity of the coagulation system. In some patients, the result may be an increased tendency for blood clots to form, and these can cause venous thrombosis, pulmonary embolism, or stroke. In women, the syndrome may also be associated with recurrent miscarriages and pregnancy complications, including preeclampsia.
The problem is that there is no single test that provides a simple and clear answer immediately. Unlike many blood tests, where one measurement is performed and a result is obtained, an investigation for lupus anticoagulant usually requires a series of coagulation tests. Sometimes there is also a need for repeat tests and various combinations between them. All this requires time, dedicated materials, and skilled laboratory staff. Even after all tests are completed, the final interpretation may be complex. Meanwhile, the waiting time for the result increases, and the direct and indirect costs of performing a full panel for each subject grow.
Search less, know more
From here we set out for a study conducted at the Hematology Laboratory at the Tel Aviv Sourasky Medical Center (Ichilov). I led the study together with Dr. Chen Hermesh, director of the coagulation testing unit, and in collaboration with Yaniv Alon and Prof. Mor Saban from Tel Aviv University. The study was based on 7,454 lupus anticoagulant investigations collected over several years at the hospital. When we examined the results, it turned out that almost 90% of the investigations ended in a negative result. In other words, in the vast majority of cases, a complex laboratory process was performed, which included on average about seven different coagulation tests, until it turned out that lupus anticoagulant was not found.
This was the moment when the practical question became the focus of the study: is it possible to identify a large part of the negative cases even before completing the entire series of tests? To find an answer, we did not look for a new device or a new test, but looked again at information that was already in our hands. We found that an important sign is hidden within two versions of a coagulation test that are already in routine use in the laboratory. Both tests measure the clotting time, but do so under different conditions. The difference between them is in the amount of phospholipids — those same fatty components that play a central role in the clotting process and are a target for pathological antibodies.
The fingerprint of the sample
Here artificial intelligence entered the picture. Using an artificial intelligence model, we found that the ratio between the results of the two tests well reflects the degree of influence of pathological antibodies on coagulation reactions that depend on phospholipids. In simple words, within the two initial coagulation tests, a kind of "fingerprint" of the sample's reaction to phospholipids was hidden. The model helped to identify this fingerprint and use it as an early screening layer, even before performing the full panel.
The goal was not to cancel existing tests, and also not to determine using the model alone who is positive. The goal was much more focused and cautious: to identify the samples where the probability of a negative result is very high. The results were encouraging. When the system classified a sample as having a high probability of being negative, the prediction was correct in about 98% of cases. But this figure does not mean that every sample can be stopped at an early stage. On the contrary: every case that is not identified with high confidence as negative continues to undergo the entire accepted array of tests. This happens both when the sample is suspected as positive and when the result is unclear.
Therefore, artificial intelligence does not replace medical judgment here and does not cancel existing tests. Its role is to help identify cases where the chance of finding lupus anticoagulant is very low, and thereby allow for the reduction of unnecessary tests without forgoing the full investigation in cases that require it. If the approach proves itself in the future as well, it may have great practical significance in the work of laboratories. It can reduce the number of complex tests, shorten waiting times, and free up manpower and resources for samples that require a more in-depth investigation.
The study also illustrates another way to think about integrating artificial intelligence into medicine. It is not always necessary to start with new and expensive equipment. Sometimes it is possible to find new value precisely within the results of routine tests that are already performed anyway, and to use the existing information in a smarter and more efficient way. The next step will be to test the approach in additional laboratories and in different testing systems, in preparation for its implementation in routine work.
The article is based on a study published in the journal Digital Medicine from the Nature group. The author is Dr. Bentzi Katz, director of the Hematology Laboratory at Ichilov Medical Center. Irena Shalev, Tania Badelbayev, Dr. Varda Deutsch, Dr. Yifat Alkalay, Dr. Ilya Kirzhner, and Dr. Roi Gat also participated in the study, and it was conducted with the assistance of Prof. Irit Avivi and Chilik Avivi.
