In our case, the effect sizes are smaller than those typically observed in neurodegenerative diseases, where the algorithm has been applied previously. Therefore, a more fine-grained choice of z-scores with a reduced maximum z-score is likely more appropriate.
In the initial implementation, we modeled only two states relative to healthy controls (HCs): unaffected (white) versus affected at 1 SD below the HC mean (red).
In the next step, we refined this approach by introducing two thresholds: 0.5 SD (red, indicating mild/moderate abnormality) and 1 SD (magenta, indicating severe abnormality). However, rather than revealing a mixture of stages and severity levels (Figure 2), the resulting patterns appeared more consistent across regions and stages.
Figure 1. Result from our data
Figure 2. Ideal results.
For these two subtypes,
Can we validate these two brain subtypes with behavior data?
What if using min-max, not z-score standardization method? that way we could capture the full spectrum of change.
A highly relevant study by Jiang et al., 2024 applied the SuStaIn algorithm and identified two distinct neurostructural subgroups by mapping the spatial and temporal trajectories of gray matter change in schizophrenia. Although the work has been published, it nevertheless received rather harsh reviewer comments.
The main problem is the SuStain algorithm with which I am very familiar. The algorithm was established primarily for use in neurodegenerative disorders. It therefore assumes a monotonic change in neuroimaging features in the direction of shrinkage. He definition of “stages” is purely arbitrary based on a z-score that is decided a priori by the investigators with limited ability to test these assumptions against some ground truth. In addition, the program can only handle very few features; we found that it works reasonably fast with 4 or 5 neuroimaging features (e.g., the lobes of the brain) but when features increase to 15 (for example) the program does not converge to any solution in addition to taking weeks to run even on supercomputing systems. Of note, the data used here are cross-sectional so the “stages” and trajectories are inferred by the degree of atrophy. I appreciate that despite its profound limitations, SuStain captures some aspect of the pathophysiology of psychosis but I am afraid that the results will join all the other studies that attempted patient classification each of which comes up with some mathematical solution without increasing our understanding of the pathophysiology of psychosis and without any clinical value in the real world.