TL;DR
This paper introduces a novel machine learning (ML) model that estimates intracranial pressure (ICP) noninvasively using pulsatile cranial expansion waveforms captured by the brain4care (B4C) system. The model achieves a mean absolute error of 3.00 mmHg, demonstrating a promising step towards a safer and more accessible method for monitoring this critical physiological parameter.
What Problem Does This Paper Address?
Intracranial pressure (ICP) is a vital indicator of brain health, and elevated ICP can lead to severe neurological damage and death. The current gold standard for ICP monitoring is invasive, requiring surgical placement of probes. This method is associated with risks, high costs, and the need for specialized personnel, limiting its widespread use, especially in resource-limited settings or emergency situations. While noninvasive methods exist, none have successfully replaced invasive techniques. This study aims to bridge this gap by developing and validating a noninvasive ICP estimation method using machine learning applied to cranial expansion waveforms.
Methods
The core of this research lies in a machine learning approach. The authors developed a model using a histogram gradient boosting (HGB) regressor, chosen for its ability to capture complex, non-linear relationships in data and its computational efficiency compared to other advanced models like TCNs and CNNs.
The process involved:
- Feature Engineering: Extracting approximately 1110 features from the B4C system's pulsatile cranial expansion waveforms (nICPW). These features capture various aspects of the waveform's morphology.
- Feature Selection: Refining these features by analyzing their correlation with invasive ICP and their calculated importance, ultimately selecting 15 key features.
- Model Training and Optimization: Employing a robust 10-fold cross-validation strategy. The model was trained on 80% of the data and tested on the remaining 20% within each fold. Hyperparameter tuning was performed using a grid search to optimize for Mean Absolute Error (MAE) and Mean Squared Error (MSE).
- Validation: A separate set of 20 patients was reserved for independent validation of the final model.
- Comparison: The developed model's performance was compared against a previously reported noninvasive method based on transcranial Doppler (TCD) to establish a baseline.
Datasets
The study utilized data from 112 neurocritical patients with acute brain injuries.
- Training and Testing: Data from 92 patients (9591 10-second windows) were used for model development (training and testing).
- Independent Validation: Data from 20 patients (2013 10-second windows) were used for final validation.
- Comparative Analysis: A subset of 13 patients from the validation group, who had concurrent B4C and TCD measurements, was used for comparing the proposed model with the TCD-based method.
The data was collected from multiple centers across Brazil, Portugal, and the United States.
Key Insights and Ideas
- Waveform Morphology as a Surrogate: The study strongly suggests that the subtle, pulsatile expansions of the skull, captured by the B4C system, contain rich information about intracranial dynamics, acting as a reliable surrogate for invasive ICP.
- ML's Power in Complexity: The success of the ML model highlights its capability to decipher complex, non-linear relationships within the nICPW that are not easily discernible through traditional analysis.
- Individualized ICP Management: The development of a noninvasive, real-time ICP estimation tool opens doors for more personalized and dynamic patient management, moving beyond rigid ICP thresholds.
- B4C System's Advantage: The B4C system's ability to derive a physiological signal directly from cranial expansion waveforms, without relying on multiple secondary parameters (like TCD or ABP), is a significant advantage for simplicity and potential accuracy.
Key Results
The developed ML model achieved promising results:
- Validation MAE: A mean absolute error (MAE) of 3.00 mmHg on the independent validation dataset.
- 95% Confidence Interval: A 95% confidence interval for ICP prediction of ±7.5 mmHg.
- Accuracy within Range: Approximately 72% of the validation estimates were within 0-4 mmHg of the invasive ICP values.
- Error Distribution: Less than 9% of the 10-second windows showed an error greater than 6 mmHg.
- Comparison with TCD: The proposed eICP model showed a stronger correlation and a narrower 95% CI (±4.17 mmHg) compared to the TCD-based method (±10.78 mmHg), which also showed no significant relationship with ICP in the tested subset.
Limitations and Future Work
The authors acknowledge several limitations and suggest future directions:
- Limited ICP Range: The dataset had a limited representation of elevated ICP (>20 mmHg), which may have affected the model's performance in these critical ranges. Future work will focus on data augmentation and training with larger, more diverse datasets.
- Patient Agitation and Artifacts: Patient agitation can reduce signal-to-noise ratio, and improper sensor placement can lead to artifacts. Supervised monitoring and the B4C system's built-in artifact detection are suggested to mitigate these issues.
- Invasive Sensor Variability: Potential measurement errors from different types of invasive ICP sensors could influence model accuracy.
- Generalizability: The model's generalizability to clinical conditions outside of acute brain injury (ABI) needs further prospective validation.
- Edge Cases: While promising, the model may not yet be sufficient for standalone clinical decision-making in all edge cases.
- Prospective Validation: The authors emphasize the need for larger prospective studies to confirm the clinical utility and generalizability of the model across diverse settings.
Summary updated