A diagnosis of left bundle branch block (LBBB) is hugely impactful for clinical treatment decisions, so it is very important to differentiate from general intraventricular conduction delay/block and/or left ventricular hypertrophy.
This paper (https://lnkd.in/gd7r52ZZ) assesses the impact of two sets of LBBB criteria from the European Society of Cardiology (ESC). It is interesting to consider that criteria can be more or less strict and demanding, corresponding to different operating points on a ROC curve. And that the two sets can lead to significant discordance in assessment.
It is clear that the 2013 criteria have much higher sensitivity. Figure 1 from the paper is outstanding in demonstrating the effect of individual criteria of varying strictness, as well as the specific LBBB morphologies that are effected. The 2021 criteria appear to reach only 30% of the 2013 criteria sensitivity.
Which set of criteria is "better", and in general how to choose an optimal operating point? Depends on the endpoint and clinical implications. It does seem like the loss of sensitivity here is very large, and indeed the new criteria seem unnecessarily highly specific. It is worth noting however that Positive Predictive Value (PPV) is often under-appreciated. A model with high PPV will earn a clinician's trust, and they will be empowered to make treatment decisions more confidently. A Data Scientist will often choose the optimal balance between sensitivity and PPV, but that is too simplistic and naïve.
Complicating this discussion is the distinction between interpreting a LBBB pattern on an ECG, and detecting a "true" LBBB with implications for Cardiac Resynchronization Therapy fitness (see https://lnkd.in/gAiqQzBa).
Epistemological limit, we're bumping into you again!