I won’t even begin to comment on the methodology of this paper - because the math exceeds my present understanding. However, it is worth highlighting that if this paper proves to be an early application of causal reasoning in the medical domain and is a valid approach, this could be a game-changer (in XXX years). As highlighted for the non-technical reader in “The Book of Why” Judea Pearl (one of the founders of modern ML), causal reasoning would allow much more generalized AI, which probably is particularly applicable to the medical space.
That said - it is worth highlighting - again, from a non-expert reader - this paper is pretty basic in its application to a medical dataset. Interestingly, the link is dead for the breast cancer data - but is findable. The dataset is from 1991, is a set of 10 features numerically rendered describing characteristics of FNA of breast mass. It is balanced (rare in real life, and significantly impacts performance of most ML algo’s using correlative methods - e.g. almost all of current ML) - with 357 benign cases and 212 malignant cases. Further, the way the data is analyzed is essentially taking 2 of the 10 features, overlapping diagnosis between the sets, and essentially the algorithm uncovers a putative causal relationship between the perimeter of the description of the cell nucleus and the “texture” and the diagnosis of malignancy. On reading the data sets description of what the data set actually is - the 10 features are rendered by an obsolete developmental computer vision system (I think from reading the paper (Teague, Cancer Cytopathology, 1997)), and actually represent an value comprised of the mean value of the cells observed, associated SE, and “worst” or largest of the observed cells. Sample image from the original paper is attached - from a modern clinician’s perspective, these images are really poor quality.
Finally - the real take home from application in medicine perspective to serve as expectation dampening - after going through all this - those that get through the paper will note that for the Breast cancer diagnosis example - it is not clear to me if the proposed causal algorithm developed by Babylon Health ACTUALLY PERFORMS BETTER THAN THE EXISTING ALGORITHM. This is not at all apparent from the final sentence of the abstract “providing a sound and complete algorithm that outperforms previous approaches on synthetic and real data”.
This led down a rabbit hole for me of trying to figure out what a MAGs is - and I still cannot ascertain what the numeric rendering of a Maximal Ancestral Graph (MAG) is, much less if 42 is better or worse than 6. I’ve attached my rabbit hole key so the rest of anyone interested in trying to interpret the value of this approach with similar background knowledge to me (e.g. limited) is able to at least wikipedia the term. 1910.11356.pdf (413 KB)