Laibaff7/cda-emr-classifier

EMR (Extended Mass Range) Binary Classifier (5-model ensemble)

A 1D-CNN ensemble that flags Extended Mass Range (EMR) events in Cassini CDA TOF mass spectra, in the high-mass region, and ranks the full unlabelled pool by EMR-likeness. Because it was trained on only 34 known EMR positives, it is an expert triage / ranking model: sort unknown spectra by the ensemble probability and manually review the top candidates.

Task & data

• Formulation: binary classification — EMR (class 1) vs. unlabelled background (class 0).
• Positives: 34 known EMR spectra. Negative pool: 130,233 spectra, of which 1,000 are subsampled for training (fixed data_seed=42), giving ~30:1 imbalance.
• Input: the spectrum cropped to channels 700-1017 (318 bins), then min-max normalised to [0,1] per spectrum. That is the only transform.

Architecture

Shallow 1D CNN (same architecture as the HMOC classifier; kept small to avoid overfitting 34 positives):

Conv1d(1->32, k=7, p=3) + BN + ReLU
Conv1d(32->64, k=5, p=2) + BN + ReLU
Conv1d(64->128, k=3, p=1) + BN + ReLU
AdaptiveAvgPool1d(1)
Linear(128->64) + ReLU + Dropout(0.4)
Linear(64->1)  -> sigmoid

Ensemble

Five independent models, one per StratifiedKFold fold, trained with seeds [42, 100, 2026, 777, 888]. At inference the mean of the 5 sigmoid outputs is the final EMR probability.

Training

• Loss: Focal Loss (alpha=0.85, gamma=2.0).
• Optimizer: AdamW (lr=1e-3, weight_decay=1e-4), batch_size=64.
• Batches: WeightedRandomSampler by inverse class frequency.
• Early stopping on validation AUPRC (patience=15); best-F1 threshold scanned per fold.

Metrics (per seed, validation fold)

Seed	AUPRC	Precision	Recall	F1	Best threshold
42	0.938	1.000	0.857	0.923	0.75
100	1.000	0.875	1.000	0.933	0.60
2026	0.860	0.857	0.857	0.857	0.50
777	0.981	1.000	0.714	0.833	0.80
888	0.878	0.545	1.000	0.706	0.75

Inference over the full 130,233 pool (ensemble mean)

• Probability distribution: mean 0.0457, p50 0.0015, p99 0.6915, max 0.9025.
• Candidates: 4,837 ≥ 0.5 · 749 ≥ 0.75 · 1 ≥ 0.9 (top: sclk 1561703929, prob 0.9025).
• Leakage audit (clean): 0 positive SCLKs appear in the inference pool; 0/100 top candidates were training negatives.

Usage

import numpy as np
from model import predict_ensemble, SEEDS, default_weights_dir

d = default_weights_dir()
paths = [f"{d}/emr_model_seed_{s}.pth" for s in SEEDS]

raw_spectra = [...]  # list of FULL raw spectra (1-D arrays); cropped to 700-1017 internally
probs = predict_ensemble(paths, raw_spectra)   # -> np.ndarray of EMR probabilities

# or, if you already have the 318-bin cropped region:
# probs = predict_ensemble(paths, cropped_spectra, already_cropped=True)

order = np.argsort(probs)[::-1]
top = [(raw_spectra[i], float(probs[i])) for i in order[:100]]   # ranked triage list

Single-model variant:

from model import load_model
net = load_model(f"{d}/emr_model_seed_100.pth")

Files

Path	Contents
model.py	Self-contained architecture, 700–1017 crop, preprocessing and ensemble inference
config.json	Architecture, training hyperparameters, per-seed metrics/thresholds
metrics.json	Per-fold validation metrics
emr_model_seed_{42,100,2026,777,888}.pth	model.state_dict() for each seed
scripts/train_emr_classifier.py	5-fold training of the ensemble
scripts/crop_unknown_data.py	Crops the unlabelled pool to channels 700–1017
scripts/infer_unknown.py	Scores the 130,233-spectrum pool with the ensemble
scripts/check_leakage.py	Leakage audit between training data and the inference pool
predictions/emr_top100.csv	Top 100 EMR candidates (original_parquet_index, sclk, emr_probability)
predictions/emr_candidates_unknown.csv	EMR probability for every spectrum in the pool
predictions/leakage_report.json	Output of the leakage audit
docs/ML_summary.md	Write-up of the approach and results

The scripts keep their original layout: they resolve paths from the workspace root that held EMR Classifier/ and Data Parquet/, so adjust ROOT / the path constants at the top of each script before rerunning. The data itself is not included.

Caveats

Trained on only 34 positives — this is a ranked triage list, not a definitive classifier. Validate top candidates before treating them as confirmed EMR events.

License

No open-source license is provided for these weights; all rights reserved. Do not reuse, redistribute, or build upon this model without the author's permission.