For the last decade we have been developing decision support tools. Decision support tools is essentially using pattern recognition to use all the amounts of health data we have on our patients to identify patients with an unmet need. What we did for the CLL Treatment-Infection Model was to start out realising that our patients newly diagnosed with CLL actually had a higher risk of severe infection than the risk of receiving CLL treatment, and for the patients who had been diagnosed with CLL and experienced a severe infection before receiving CLL treatment they had a one month mortality rate of 10%.
So we had a clear unmet need. We had not been aware of this in the clinic because we would not see these patients in the haematology clinic, they would be admitted to a regular emergency room upon the infection. So we set out to find a way to identify these patients at the time of diagnosis and we used all the available data and developed the CLL Treatment-Infection Model, CLL TIM, which we published a few years ago.
To validate that algorithm we also used an external German CLL trial and we have also now recently had colleagues from the Mayo Clinic validating this algorithm. But what is most important is, as if it was a new drug, we also need to validate prospectively that the algorithm works and that’s what we have been presenting here – the prospective validation in the PreVent-ACaLL clinical trial.
What were the key results?
What we reported this EHA meeting is the results from the non-randomised part of the trial. We pre-screened a bit more than 500 patients and the algorithm would assess them as high risk or low risk in terms of the event of infection or CLL treatment in the first two years from diagnosis. What we report here is for the four groups of patients because in addition to the high risk/low risk assessment we also assessed the confidence that the algorithm would have in the estimate of whether the patient would be high risk or low risk. So we had a high risk/high confidence group, high risk/low confidence group, low risk/low confidence group and low risk/high confidence group.
For the trial it was defined that only the high risk/high confidence group, the patients being at the highest risk of a severe infection or CLL treatment and the algorithm assessing this with a high confidence, that would be eligible for the trial for the randomised part of the trial. For those patients, half of the patients, actually a bit more than half of the patients, eventually were not randomised due to patient choice or due to assessment of the investigator.
We report on this half of the high risk/high confidence patients and the three other groups to see prospectively if the algorithm could actually in a change of time, it was trained before 2019 and the trial was enrolling from 2020 onwards, so it was a change of time, it was a change of geography, meaning that we also included Swedish and Dutch sites, and it was a change of medical environment, both in terms of other hospitals, other medical cultures, but also being during the COVID pandemic.
What we see in the trial here is that the algorithm is robust. It can still differentiate or identify the patients being at high risk of severe infection or CLL treatment in the high-risk groups as compared to the low-risk groups. We can also demonstrate that the confidence estimate is still valid and robust in this change of time and change of location.
Obviously the next step now would be to test if the algorithm could actually identify these patients and the intervention would have an impact in terms of lowering the risk of CLL treatment and lowering the risk of severe infections. That’s what we’re waiting for, is the next part of the trial to report on the randomised patients.
