NEXVARA Independent Research · Hong Kong

AI that learns
this patient, not every patient.

Population-trained models average away the individual variation that defines critical care. We build locally-adaptive systems that model each patient's physiology from the ground up — and improve as they grow.

60/65
Subjects with positive separation vs population baseline
489
ICU patients in continual learning validation
p .0014
Continual learning trend significance · MIMIC-III
4
Independent datasets · neonatal through adult ICU
Three-paper clinical arc

The argument, in sequence.

Each paper closes the gap left by the previous. Together they constitute a complete, deployable system.

Paper 01
Live on SSRN
Per-Individual Generative Modelling for Physiological Early Warning

The why. Per-individual autoencoders outperform population models on anomaly detection across four independent clinical datasets spanning neonatal EEG, preterm cardiac, paediatric epilepsy, and adult ICU. Introduces reconstruction error as free energy.

60/65 subjects positive separation · 7.2× over population baseline · Gradual / Acute phenotype emergent
doi:10.2139/ssrn.6756441
Paper 02
Submission pending
Compatibility-Filtered Warm-Start Personalisation

The how to start. Reconstruction cross-error identifies compatible donors prospectively. Warm-starting from their weights outperforms both cold-start and uniform initialisation. Introduces prior contamination as a structural failure mode of averaging.

p = 0.0156 filtered > uniform · 5% and 20% calibration · PICS neonatal n=9
Paper 03
Target: npj Digital Medicine
Self-Routing Continual Learning for ICU Monitoring

The how to run. Early fine-tuning loss dynamics detect compatibility without outcome labels. Unsupervised clustering recovers clinically meaningful phenotypes. The pool improves autonomously — no central retraining required.

k-means ARI = 0.60 · slope −0.0015/patient, r = −0.462, p = 0.0014 · MIMIC-III n=489
Architecture

A system that knows
you are not average.

No global model to retrain. No population priors contaminating individual predictions. Five components, one coherent loop.

🧬

Patient Onboarding

Each new patient's autoencoder is trained on their own physiological signal. Compatibility against the pool is assessed via cross-reconstruction error before any prior is selected.

Self-Routing

Early fine-tuning loss dynamics (epochs 1–10) detect compatibility without event labels. Compatible patients warm-route from donor priors; isolated patients receive cold-start initialisation.

📈

Continual Learning

Every new patient enriches the compatibility geometry. Warm-routing advantage increases significantly as pool density grows — without any central retraining step.

🔬

Anomaly Monitoring

Real-time reconstruction error surfaces the model's surprise at each incoming signal segment — deviation from that individual's learned baseline, not a population average.

"The same architecture that predicts a neonatal seizure before it happens can predict any system's failure, optimum, or next evolution — before it happens."

The ICU is the hardest test case for locally-adaptive AI. Inter-individual physiological variability is clinically non-negotiable. Ground truth is unambiguous. The cost of wrong priors is legible.

Demonstrating the principle here — where it is most difficult to dismiss — establishes it for every domain where individual variation matters: material degradation, supply chain, human behaviour, scientific discovery.

About

NEXVARA
Independent Research

N
Noelle Xiao
Founder · Hong Kong
Neuroscience Locally-adaptive AI ICU monitoring Continual learning Free energy principle

NEXVARA is an independent research company founded on the thesis that neuromorphic, locally-adaptive AI outperforms population-trained models in efficiency, accuracy, and sustainability — not as a niche application, but as a general principle.

The dominant paradigm in clinical AI rests on the assumption that individual patients are sufficiently similar to benefit from population-level priors. In the intensive care unit, this assumption fails visibly. Two patients with identical admission diagnoses may follow entirely different physiological trajectories. A model trained on ten thousand admissions will be well-calibrated on average and systematically wrong for this patient in this deterioration window.

This is not a calibration problem solvable by larger datasets or more sophisticated architectures. It is a structural mismatch between how population-trained models represent knowledge and how individual patients actually vary. NEXVARA's answer is not to scale the population model — it is to replace it.

We conduct this research without institutional affiliation by design. The independence is the point: work accountable to evidence and argument, not to the political economies of academic departments or the commercial interests of established players. The governing principle is temporary custodianship — knowledge held carefully, released well.

Research Collaboration
Let's build the next experiment together.

NEXVARA is building toward clinical deployment and actively seeking research collaborators with access to prospective ICU data. If the work resonates, reach out.