Harshith.
←Back to work
Neurava
AI/ML Algorithms Engineer · Jan 2026 — present

Sleep Staging
— biosignals · wearables · clinical validation

Sleep-staging and biosignal ML for a clinical wearable — building evaluation pipelines that compare deep sequence models and classical baselines against polysomnography ground truth. Proprietary work; details limited by NDA.

Problem

Clinical sleep assessment traditionally depends on polysomnography — multi-channel PSG recorded in a lab and scored by experts. Wearable devices need to recover useful sleep signals from far fewer sensors, without sacrificing the rigor that clinicians and regulators expect.

The core challenge is bridging that gap: how do you train and validate models on limited wearable inputs when the only trustworthy reference is full PSG?

Approach

At Neurava I work on the algorithms side of this problem — designing pipelines that connect reference-grade sleep annotations to wearable sensor streams, then training and benchmarking models under held-out subject evaluation.

The work spans several layers, described here only at a high level:

  • Reference signal processing — generating sleep-stage labels from multi-channel physiological recordings using established staging tooling and internal validation workflows.
  • Wearable modeling — training sequence and classical models on paired sensor data, with per-subject normalization, augmentation, and class-imbalance handling suited to overnight recordings.
  • Baseline comparison — benchmarking deep models against actigraphy-style and gradient-boosted baselines so improvements are measurable, not assumed.
  • Reproducible experimentation — versioned training configs, experiment logging, and subject-level evaluation splits so results are traceable across dataset and label revisions.

Specific architectures, hyperparameters, performance numbers, and deployment details are confidential Neurava intellectual property and are not disclosed here.

Scope

This is active, proprietary work. I can speak to the general problem domain — PSG-grounded validation, wearable IMU modeling, and rigorous ML experimentation — in conversation, but cannot share internal repository structure, model specifications, clinical dataset details, or product-specific sensor configurations on a public site.

No public repository or demo for this project. Interested parties can reach out directly; anything beyond the summary above requires Neurava approval.