Development and internal validation of a CSF-free machine-learning model to distinguish immune checkpoint inhibitor-induced encephalitis from HSV-1 and anti-LGI1 encephalitis

Abstract Background Immune checkpoint inhibitor-induced encephalitis (ICI-iE) is difficult to distinguish from herpes simplex virus type 1 (HSV-1) and anti-LGI1 encephalitis, and its work-up includes invasive cerebrospinal fluid (CSF) testing. We evaluated whether a model excluding CSF features could achieve discrimination similar to a CSF-inclusive model and support triage before lumbar puncture. Methods We reanalysed 51 complete cases (ICI-iE, 19; HSV-1, 20; anti-LGI1, 12) from a publicly available retrospective dataset combining registry-based ICI-iE cases with single-centre HSV-1 and anti-LGI1 comparators. Random-forest models with 29 features (including CSF) or 25 features (excluding CSF) were evaluated by repeated nested cross-validation (5 outer folds × 20 repeats). We assessed discrimination, calibration, decision curves, out-of-fold TreeSHAP values, and a hypothetical high-specificity rule-in strategy across assumed prevalences. Results Mean areas under the receiver operating characteristic curve (AUCs) across repeats were 0.894 with CSF and 0.902 without CSF. On the repeat-averaged out-of-fold predictions, AUCs were 0.911 with CSF and 0.908 without CSF, and an approximate DeLong comparison showed no significant difference (ΔAUC [with CSF minus without CSF], + 0.003; 95% CI, − 0.033 to + 0.040; p = 0.86). Encephalitis-typical magnetic resonance imaging (MRI) findings and C-reactive protein were the leading contributors. In an exploratory sensitivity analysis, additionally excluding MRI features from the CSF-free model yielded a mean cross-validated AUC of 0.769. At specificity of 0.967–0.973, sensitivity was 0.574–0.597. Under the hypothetical rule-in strategy, the true-positive classification proportion was 21%–22% at the cohort prevalence (37.3%) but 3%–6% at 5%–10% prevalence. False-positive classification remained 2–3% of patients across the evaluated prevalence scenarios. Calibration was poor. Conclusions In this small internally validated dataset, excluding CSF features did not materially lower the point estimate of discrimination, whereas discrimination depended substantially, but not exclusively, on MRI findings. The CSF-free model may support triage while confirmatory testing is pending but should not replace CSF testing to exclude HSV-1 encephalitis. External validation in standardised, representative cohorts is required, particularly given potential confounding by data source.