Research notes.

Non-invasive ICP estimation in pediatric neurocritical care

2025 – Present · ongoing research · PyTorch · LSTM, TCN, Transformer, Hybrid · LOPO cross-validation

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 thesis asks a simple question: how close can we get to that number from waveforms a bedside monitor already records?

The setup uses two channels, arterial blood pressure (ABP) and cerebral blood flow velocity (CBFV), resampled to 125 Hz and chopped into 10-second windows with a 2-second stride. The training data is PhysioNet's pediatric neurocritical-care waveform release (12 patients, around 18,500 windows), with PhysioNet's CHARIS used as an adult cross-dataset validation set.

Validation is leave-one-patient-out, so each fold tests on a patient the model has never seen. That cuts down the false confidence you get from random splits, where adjacent windows from the same patient leak between train and test. Loss is Huber, learning rate is cosine-scheduled, models train for up to 100 epochs with patience-15 early stopping.

Models tried

• LSTM. Bidirectional, 2 layers, 64 hidden units, with attention. Currently the strongest single model.
• TCN. Temporal convolutional network with dilated causal convolutions; channels [32, 64, 64, 128], kernel 7.
• Transformer. Encoder with learnable positional encoding, 4 heads, 3 layers.
• Hybrid. CNN feature extractor into a Transformer, with a small autoregulation module on top.

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)
• Folds: 12 leave-one-patient-out, 18,485 total windows

Per-patient errors range widely. Patient 18 lands at MAE 0.65; patient 21 sits at MAE 5.26. The next round of work focuses on the wide-error patients, with a class-balanced training schedule (the results_balanced_inverse_freq_boost branch) and morphology-aware features at the 60-beat level.

Reference points

• Fanelli et al., Journal of Neurosurgery: Pediatrics, 2019. The dataset paper for the pediatric release.
• Pseudo-Bayesian Model-Based Noninvasive Intracranial Pressure Estimation and Tracking. Model-based baseline.
• A Spectral Approach to Model-Based Noninvasive Intracranial Pressure Estimation. The other model-based baseline; both run on similar inputs and are the bar a learned model has to clear to be worth the bother.

Supervisor: Prof. Zubaer Ibna Mannan, Smart Computing, Kyungdong University. PhysioNet credentialed access to both datasets.

Temporal heart-defect prediction (CheXchoNet)

2023 – 2024 · senior thesis · PyTorch · CNN with temporal metadata · CheXchoNet

Predicting heart defects from chest X-rays alongside the temporal metadata that comes with them, rather than treating each frame in isolation. A CNN trained over the CheXchoNet release reached AUC of 0.79 to 0.84 in cross-validation. The approach pointed me towards waveform-level signals, which is where the ICP work picks up.

Supervisor: Prof. Zubaer Ibna Mannan, Smart Computing, Kyungdong University.

Cardiovascular disease prediction

2023 · Scikit-learn · PyTorch · Cleveland Clinic dataset

First real ML research project. A binary classifier over the Cleveland Clinic dataset, predicting cardiovascular disease from a small set of patient features. The interesting work was less the model and more the surrounding craft: train/test splitting, feature engineering, calibration, and learning to write evaluation that doesn't lie to you. Final model reached 96% test accuracy.