Research Question

The Core Question

How do preprocessing choices (outlier detection → imputation) affect downstream prediction quality when using handcrafted physiological features for glaucoma screening?

What This Means

We FIX the Classifier

CatBoost is established as the best classifier. We don't compare classifiers—that's not the research question.

We VARY the Preprocessing

We systematically test combinations of:

• 11 outlier detection methods (foundation models vs traditional)
• 7 imputation methods (deep learning vs classical)

We MEASURE Downstream Effects

Not just AUROC, but ALL STRATOS-compliant metrics:

• Discrimination (AUROC)
• Calibration (slope, intercept, O:E ratio)
• Overall (Brier, Scaled Brier)
• Clinical utility (Net Benefit, DCA)

What This Is NOT About

❌ NOT This	✅ This Instead
Comparing classifiers	Fix classifier, vary preprocessing
Maximizing AUROC	Measure all STRATOS metrics
Generic ML benchmarking	Preprocessing sensitivity analysis
Foundation model features	Handcrafted physiological features

Why Foundation Models for Preprocessing?

Foundation models (MOMENT, UniTS, TimesNet) are generic time-series models trained on large datasets. We ask:

1. Can they detect artifacts as well as human experts?
2. Can they reconstruct missing signal segments accurately?
3. Do they reduce error propagation compared to traditional methods?

Key Insight: Embeddings Underperform

Foundation model embeddings (used as features) underperform handcrafted features by 9 percentage points (0.740 vs 0.830 AUROC).

Why? Generic embeddings don't capture domain-specific PLR physiology. Handcrafted features encode expert knowledge about glaucoma biomarkers.

Data Provenance

Dataset	N	Source
Najjar et al. 2023	322	Original study
Our classification subset	208	152 control + 56 glaucoma
Our preprocessing subset	507	All with ground truth masks

Do Not Compare Directly

Our AUROC (0.913) cannot be compared to Najjar's (0.94) due to different subsets and different goals.