In arXiv:2311.10798, 12-month PH is presented as a benchmark task, with AUROC results reported in Table 5. In https://github.com/som-shahlab/INSPECT_public/blob/main/ehr/2_generate_labels_and_features.py#L283-L293, the intended SourceCodeLabeler path for 12_month_PH is commented out (presumably because the source codes were not released due to PHI concerns). Instead, the script reads 12_month_PH directly from the cohort CSV using True/False/Censored labels.
However, run_all_ehr.py still runs PH through GBM/MOTOR, so the training pipeline functions correctly, but it appears to rely on precomputed cohort labels rather than the EHR phenotyping approach described in the paper (ICD codes + Appendix D).
I was wondering whether the PH results reported in the paper were generated using SourceCodeLabeler or the cohort CSV columns.
In arXiv:2311.10798, 12-month PH is presented as a benchmark task, with AUROC results reported in Table 5. In https://github.com/som-shahlab/INSPECT_public/blob/main/ehr/2_generate_labels_and_features.py#L283-L293, the intended SourceCodeLabeler path for 12_month_PH is commented out (presumably because the source codes were not released due to PHI concerns). Instead, the script reads 12_month_PH directly from the cohort CSV using True/False/Censored labels.
However, run_all_ehr.py still runs PH through GBM/MOTOR, so the training pipeline functions correctly, but it appears to rely on precomputed cohort labels rather than the EHR phenotyping approach described in the paper (ICD codes + Appendix D).
I was wondering whether the PH results reported in the paper were generated using SourceCodeLabeler or the cohort CSV columns.