Nexvara Early Warning

Per-Individual Physiological Anomaly Detection · Preprint 2026

A neonatal patient's baseline is not the population average. It is their own.

---

What this is

Standard clinical monitoring compares a patient's signals against population thresholds — the same alarm limits for every infant, every adult, every ICU patient. This works until it doesn't: a preterm infant with an unusually low baseline heart rate triggers constant false alarms, while a patient whose "normal" sits at the edge of population norms deteriorates silently within range.

Nexvara Early Warning takes a different approach. A lightweight autoencoder is trained on each individual patient's own clean baseline — the first hours of stable signal after admission. From that point forward, the model scores every new window against that patient's learned normal. Deviation from the individual baseline, not the population mean, is the anomaly signal.

The result: the model knows this infant, not infants in general.

---

Results

Validated across four independent clinical datasets spanning the full lifespan:

Dataset	Population	Modality	Patients detected	Mean separation	Lead time
Helsinki Neonatal EEG	Neonatal	EEG	10/10	3.35σ	~10s before seizure
PICS Preterm	Preterm infant	HR / RR	9/9	3.37σ	30s (p<0.0001, n=581)
CHB-MIT Paediatric EEG	Paediatric	EEG	9/9	7.22σ	p=0.022 @ −10s
MIMIC-IV Adult ICU	Adult ICU	HR / SpO₂ / RR	32/37	3.66σ	~2 hours (p=0.008)

vs population model baseline: 0.83σ — a 7.2× improvement in separation on CHB-MIT.

Cross-subject control: applying one patient's model to another patient's data collapses separation by 622×, confirming the per-individual specificity of the approach.

---

Clinical significance

• 30 seconds before bradycardia/apnea onset in preterm infants — enough time to pre-position staff and prepare intervention
• ~2 hours before vasopressor initiation in adult ICU — a window that does not exist in current monitoring
• Zero false positives on the individual's own stable baseline in the majority of patients
• Identifies two distinct deterioration phenotypes in MIMIC-IV: Gradual (detectable ~2h ahead) and Acute (rapid onset, different clinical pathway)

---

Architecture

Patient admitted
      │
      ▼
┌─────────────────────────┐
│  Baseline window        │  First stable hours of signal
│  (individual-specific)  │
└────────────┬────────────┘
             │  train
             ▼
┌─────────────────────────┐
│  Per-individual         │  Lightweight convolutional
│  autoencoder            │  autoencoder, ~50k parameters
└────────────┬────────────┘
             │  reconstruct
             ▼
┌─────────────────────────┐
│  Anomaly score          │  Reconstruction error vs
│  (continuous)           │  individual baseline distribution
└────────────┬────────────┘
             │
             ▼
        Early warning

One model per patient. No shared weights between patients. No population reference required after training.

---

Demo

Interactive clinical dashboard — switch between datasets, patients, and signal views.

pip install streamlit matplotlib numpy pandas
streamlit run demo/nexvara_demo.py

Shows:

• Live signal + anomaly score with lead time window highlighted
• Per-individual vs population model comparison (false alarm rate contrast)
• Cross-dataset results summary with real preprint numbers

---

Repository structure

nexvara-early-warning/
├── README.md
├── requirements.txt
├── demo/
│   └── nexvara_demo.py        # Streamlit clinical dashboard
├── figures/                   # Key result figures from preprint
│   ├── separation_comparison.png
│   ├── lead_time_pics.png
│   └── population_vs_individual.png
└── notebooks/
    └── per_individual_baseline.ipynb   # Core method walkthrough

---

Preprint

Per-Individual Generative Modelling for Physiological Early Warning Across the Lifespan Nexvara Research, 2026 MedRxiv: [DOI link — update on posting]

---

Datasets used

• Helsinki University Hospital Neonatal EEG
• PICS Preterm Infant Cardiorespiratory
• CHB-MIT Scalp EEG
• MIMIC-IV Clinical Database

All datasets are publicly available via PhysioNet or Zenodo. No patient data is included in this repository.

---

Contact

Nexvara Research For research collaboration enquiries: [your email]

Registered in Hong Kong · Research focus: per-individual physiological AI

---

Citation

If you use this work, please cite:

@article{nexvara2026,
  title   = {Per-Individual Generative Modelling for Physiological Early Warning Across the Lifespan},
  author  = {Noelle Xiao},
  journal = {medRxiv},
  year    = {2026},
  doi     = {[DOI — update on posting]}
}

---

_{This repository accompanies a preprint that has not yet undergone peer review. Results should be interpreted accordingly.}