Widi Nugroho

1. Hemoglobin Estimation from Facial Videos (rPPG / MCD-rPPG)

Non-invasive estimation of hemoglobin levels from facial videos using remote photoplethysmography (rPPG). Passive facial videos are processed into physiological features, which are then mapped to hemoglobin values — an approach that requires no specialized sensors.

Best result (attention + demographics, subject-level / anti-leakage evaluation): MAE 0.934 g/dL · R² 0.470 · Pearson 0.732

Finding reported honestly: adding modalities (static color + rPPG waveform) did not beat demographics alone — R² 0.271 vs 0.470. This negative result is intentionally kept in the research documentation rather than hidden.

Tools: Python · PyTorch · 3D-CNN · rPPG · Computer Vision

2. Deconstructing Accuracy Claims: A Data Leakage Finding (replication of a PPG→Hb paper)

Replication of a paper claiming MRE 2.46% and R² 0.97 for hemoglobin estimation from PPG. Re-evaluated under a correct protocol — subject-level split on a dataset of 68 subjects, 816 rows.

Valid subject-level results: MRE ~10.5–13% · R² ~0.14. Numbers close to the claims (MRE 2.71%, R² 0.84) only appear when the evaluation uses a row-level split — a hallmark of data leakage.

This is the core value proposition: critical thinking and methodological discipline in evaluation. Accuracy claims in the literature cannot be trusted until their evaluation protocol is examined.

Tools: Python · scikit-learn · PyTorch · Statistics

3. Replicating a 3D-CNN Architecture (MDPI paper)

Replication of a 3D-CNN architecture for medical image classification, evaluated with subject-independent 5-fold cross-validation.

Critical note: the official architecture in the paper's repository turns out to differ from the description inside the paper itself (end-to-end waveform vs statistical features). As a result, the replication results are not equivalent to the publication's claims — small differences in code lead to large differences in results.

Tools: Python · PyTorch