Analytical Model Review

Model Review — NTX-101

Primary and benchmark predictive model documentation supporting the Phase II responder analysis.

Integrated report Program overview Data lineage Model review Phase I Phase II Phase III

Model Card — Phase II

A concise, sponsor-style model card describing use, inputs, benchmarking, and performance metrics for the Phase II responder model.

Use Case

Predict Phase II Week 24 responder status for exploratory program gating support.

Primary Model

logistic_regression

Inputs

Type	Variables
Numeric	age, baseline_cognitive_score, chg_week12, early_safety_discontinuation_before_week12
Categorical	sex, race, baseline_severity_group, treatment_arm

Exclusions / Leakage Controls

No post-Week-24 features used.
Outcome-defining fields at Week 24 excluded from predictors.
Training features documented explicitly in this card.

Benchmarking

Conservative benchmark set includes logistic regression, elastic net, and random forest. Logistic regression remains primary for interpretability and sponsor-style transparency.

Threshold Logic

Primary classification threshold: 0.495

Threshold selected from the training ROC curve to preserve clinically useful sensitivity while improving specificity over the default 0.50 cutoff.

Primary Model Metrics

Threshold	0.4949
ROC AUC	0.868
PR AUC	0.845
Brier	0.145
Sensitivity	0.806
Specificity	0.75
PPV	0.758
NPV	0.800
F1	0.781
N (test)	63
Responder diff (pp)	19.073
Gate decision	GO
Primary endpoint p-value	0.0019
Threshold rationale	Training ROC threshold maximizing specificity with sensitivity ≥ 0.78

Limitations

Findings should be interpreted within the documented source-mix and continuity assumptions used for this integrated review.
Predictive performance may vary if endpoint prevalence or follow-up assumptions materially change.