Non-invasive ICP estimation in pediatric neurocritical care

Intracranial pressure is one of the more consequential numbers in a neurocritical-care unit, and getting it usually means a catheter through the skull. The work asks a simple question: how close can we get to that number from waveforms a bedside monitor already records?

Two channels, ABP and CBFV, resampled to 125 Hz, 10-second windows with a 2-second stride. Trained on PhysioNet's pediatric neurocritical-care waveform release (12 patients, ~18,500 windows), with PhysioNet CHARIS as the adult cross-dataset validation set. Validation is leave-one-patient-out.

Best LSTM (LOPO): MAE 2.92 mmHg, median AE 1.97 mmHg, RMSE 4.03 mmHg, bias −1.46 mmHg.

Setup

Two-channel input: arterial blood pressure (ABP) and cerebral blood flow velocity (CBFV), both resampled to 125 Hz and split into 10-second windows with a 2-second stride. The training set is PhysioNet's pediatric neurocritical-care waveform release (12 patients, ~18,500 windows). PhysioNet CHARIS (adult ICP) is held out as a cross-dataset validation set.

Models

• Bidirectional LSTM with attention (current best). 2 layers, 64 hidden units, attention pooling over time, dropout 0.2.
• TCN with dilated causal convolutions; channels [32, 64, 64, 128], kernel 7.
• Transformer encoder with learnable positional encoding, 4 heads, 3 layers, d_model 64.
• Hybrid: CNN feature extractor → Transformer → small autoregulation module on top.

Validation

Leave-one-patient-out (LOPO) across all 12 patients. Loss is Huber, learning rate is cosine-scheduled, max 100 epochs with patience-15 early stopping.

Best results so far (LSTM, LOPO)

• Mean absolute error: 2.92 mmHg
• Median absolute error: 1.97 mmHg
• Root mean squared error: 4.03 mmHg
• Bias: −1.46 mmHg (sde 3.76)

What's next

Per-patient errors range widely (patient 18 is at MAE 0.65; patient 21 at MAE 5.26). The next round of work focuses on the wide-error patients — class-balanced training with inverse-frequency weighting, plus morphology-aware features at the 60-beat level.

Reference baselines

• Fanelli et al., Journal of Neurosurgery: Pediatrics, 2019 — the dataset paper for the pediatric release.
• Pseudo-Bayesian Model-Based Noninvasive ICP Estimation and Tracking.
• A Spectral Approach to Model-Based Noninvasive ICP Estimation.