Evidence status
Findings and interpretation
This page separates source-cohort evidence, newly computed PenuX feature importance and targets that still require validation.
Research use only. Published benchmark values are not relabelled as prospective PenuX clinical performance.
Current findings
| Finding | Value | Interpretation |
|---|---|---|
| Multi-ML source records | 1,289 | 204 SAP (15.8%) after correct target interpretation. |
| Raw target coding | 0=SAP, 1=non-SAP | The source model code explicitly inverts the raw target before modeling. PenuX now records this correction. |
| Development split used for gain export | 1,031 | 258 records held out and not used to compute feature-gain weights. |
| Class weight | scale_pos_weight ≈ 5.33 | Reflects the minority SAP class after target normalization. |
| Top XGBoost gain feature | Ca ≈ 7.21% | Largest normalized gain in the current all-feature development model. |
| Next gain features | α-HBDH 4.59%, CRP 4.08%, WBC 3.28% | Consistent with the source paper's emphasis on calcium, inflammatory markers and α-HBDH. |
| Published XGBoost discrimination | AUROC ≈ 0.92 | Source-cohort literature benchmark, not yet a prospective PenuX result. |
| PenuX watch target | Sensitivity ≥98% | Must be tested using a development-locked threshold and external validation. |
| Clinical deployment | Not established | Retrospective/open-data results do not validate bedside use. |
Top learned gain weights
| Rank | Feature | Gain weight |
|---|---|---|
| Loading… | ||
Important.
Gain is unsigned feature importance. It is not a linear coefficient and does not tell you that doubling a laboratory value doubles risk.