I have recently run into issues regarding PixInsight Sub Scoring vs. DSS, with some cases with PI rating very bad subs very highly, while DSS scored them low.
These were subs with worse light pollution (generally as targets rotated into city light dome), with high Sky Background brightness and lower number of stars. For some strange reason PI gave these subs a high SNR rating, when it is visually obvious that they had low SNR (very poor contrast and fewer stars). PI did recognize that these subs had much fewer stars than the other subs, however. My response is to delete these bad subs, because PI would erroneously overweight these, leading to a worse integration result.
This strange behavior is in addition to another strange behavior that I notice a couple of weeks ago with PI giving tremendously high SNR ratings to subs with bad jet streaks. In this case only 3 subs with jet streaks out of more than 200 gave integration results that were TOTALLY useless, but after remove the 3 offending subs, the results were good. (In this case, these bad subs were actually fair good except for the jet streaks.)
This reinforces my opinion, that you should use BOTH PI and DSS when determining which subs to delete before stacking. You should also visually investigate (I prefer DSS for this) the subs whose SNR spikes above the rest just as much as the subs that get bad scores.
PI SubFrameSel does provide some value added when used, as it gives an Eccentricity measurement, that DSS does not. BUT, it seems that PI's SNR measurement can be erratic. A sub that PI rates highly but receives a poor DSS score is probably a bad sub. You should use both, even though it slows the process.
I am using the latest PI 18.104.22.168.