Machine Learning may assist MSB in identifying ideal treatment populations
As MSB and Novartis pore over biomarker data from the COVID-19 trial, they will no doubt look for patient population and temporal trends to identify ideal candidates for Rem-L treatment. This needs to be established to set recruitment criteria for a confirmatory trial, or alternatively for a proposed label / triage criteria to be discussed with the FDA for an EUA.
Meanwhile, doctors and researchers across the world are urgently looking at data to answer the question: which patients are most likely to die of COVID-19? There are now numerous papers available which predict the prognosis of COVID-19 given numerous factors, some of which are using Machine Learning (ML) to identify those most vulnerable in terms of various patient characteristics and biometric data.
One large scale European study of 5594 patients (Jimenez-Solem et al, Feb 2021) applied ML to predict adverse outcomes based on demographic traits, comorbidities, temporal features and in-hospital tests across the stages of management of the disease. This such as diagnosis, admission, Pre-ICU, ICU/ventilator treatment and death. This study may be particularly relevant for MSB's analysis because it identifies how predictive factors are different at different stages of the disease. The following discusses some observations coming out of this study and how they may be relevant to MSB's trial.
Age shown as an overwhelmingly strong predictor of death This is self-evident / something we already knew about. However, this data shows just how overwhelming this is. Regardless of the stage of the disease or treatment, age is by far and away the strongest predictor of death.
This is relevant for MSB because it provides strong, objective, external data to justify their statistical assessment of the subgroup under 65yo. i.e. it supports that MSB are not "data-mining" randomly to "get lucky" with a particular subset. They are making exclusions for a reason based on external data.
Adding temporal and in-hospital test data (in addition to demographic data) very strongly improves ability to predict adverse outcomes
This study presents that when temporal and in-hospital test data are available and added to the ML model (i.e. in addition to a base prediction model based only on demographic data), the predictive power strongly increases. For example, at time of hospital admission, the hospital tests significantly increase prediction accuracy (p<0.01). This is indicated in the table at this link.
This is relevant to MSB as a further trial could recruit in a targeted manner not just based on demographics (e.g. under 65s) but also with the range of in-hospital tests which are predictive of mortality.
The in-hospital test of CRP is a highly significant predictor of mortalityThe graph below shows how the relative importance of factors changes once in-hospital test factors are reflected in the model in addition to demographic. Note that C-Reactive Protein (CRP), which is an inflammation biomarker, is highly significant.
The study provides a table showing median CRP for non-survivors of 79.5mg/l is significantly more than that for survivors of 59.5mg/L (p<0.001). I have discussed in a prior post how MSB's criteria for CRP of 40mg/L may well have been set considerably too low (allowing too many survivors in the control group to survive and thus make it harder to show Rem-L's efficacy). A future trial recruit based on a higher CRP threshold.
The only biomarkers which showed to be more significant were blood urea nitrogen and creatinine are indicators of kidney disfunction. IMHO, the strong significance of CRP is exciting and bodes well in terms of Rem-L's MOA being to suppress inflammation.
Using factors to predict adverse patient outcomes is stronger earlier in patient management
The nuance that I found so interesting and exciting with this study is that the ability of a model to predict adverse outcomes (particularly death) using demographic, temporal and hospital test data, is significantly stronger at hospital admission compared to later stages of management. In particular, the data showed that the power of a model to predict mortality for ventilated patients is actually less compared to admission or ICU entry.
This may be a function of ever-decreasing sample sizes as fewer patients progress to further stages of treatment. However, it may also be due to the speed or variability of the disease. The report author comments on the very rapid changes in inflammation that drivers and prognostic markers of adverse outcomes represent a dynamic field affected by the patient’s current point on the disease trajectory. The author also observes significant variations in national factors such as isolation policies and triage for ICU and mechanical ventilation, population demographics etc. may impact on results.
This is important due as MSB's trial was primarily based on patients who had progressed to ventilation, when inflammation may have subsided. A confirmatory trial might show stronger efficacy for Rem-L if recruitment criteria are based on the model's predictive factors at admission and not (always) waiting for ventilation.
Conclusion
Now more so than ever, MSB have data both within their own COVID trial, and elsewhere, to establish and demonstrate efficacy for Rem-L's ideal patient population. Whilst we keenly await an update from SI in the very near future, I would not be surprised if takes quite a while longer yet as MSB and Novartis make comparisons not only with the data generated from this trial, but from data and models produced elsewhere.
Some other resources
For those who don't like to read a lot, a concise summary of predictive factors is provided by Doctor Mike Hansen on youtube:
(20min delay)