PIs: Tamas Budavari (Whiting School of Engineering) & Rohan Mathur (School of Medicine)
Patients diagnosed with intracranial tumors, after consultation with a neurosurgeon, present to the hospital at a scheduled date for surgical resection of their intracranial tumors. Post-operatively, they are routinely admitted to the Neurosciences Critical Care Unit (NCCU) for continuous monitoring. The problem that motivates our project is that only a small unknown percentage of those patients truly develop medical conditions that require care that can only be provided in the NCCU, whereas most patients would be able to receive necessary care in less resourceintensive facilities in the hospital. Identifying those patients would allow for a more effective utilization of limited resources (i.e., total of 24 beds in NCCU) and better care for those who truly need it, and hence save more lives. Financially, today this is a multimillion-dollar issue even just at the Johns Hopkins Hospital!
We aim to identify the need for critical care of recovering patients by using recent advancement in AI/ML by utilizing the Precision Medicine Center of Excellence in Neurocritical Care (PMCoE-NCC) Data Repository that contains clinical electronic medical records or EMR data, physiological and imaging data from all patients who have been admitted to the NCCU at the Johns Hopkins Hospital.
Statistical classification methods are mature approaches that consider uncertainty and provide well-calibrated probabilistic results. That said, the recent success of deep learning cannot be overlooked, but their blind application is simply too dangerous! We will combine the power of statistics and deep learning approaches without their drawbacks.
The usual statistical methods suffer from naive feature extraction processes, which limit their applicability. For example, image classification is traditionally performed on simple summaries of segmented regions, such as their average colors or brightness, etc. Deep learning, however, is arguably so successful because it starts with the original raw data, e.g., the image itself, and learns to extract optimal features along the way. The problems begin when a fully connected neural network makes blackbox predictions at the end of the feed-forward network architecture.
We will design and train neural networks to achieve our objective in a way that the extracted features can also be used in proper probabilistic classification approaches. Using appropriate network architectures, we will identify the layers where important image properties are encoded and use them for statistical analysis. Given that the functional form of a neural network is analytically known, we can even propagate uncertainties in images to the optimal features, which can then be used in probabilistic classification, potentially even combining with other traditional features.
Specific Aim 1. Using the PMCoE-NCC database, we will first identify those patients, admitted after elective brain tumor surgery, that truly needed treatment that can only be administered at the NCCU to separate them from those who could have received care in less resource-intensive facilities.
Specific Aim 2. To establish a baseline, we will first use the usual traditional features available for all patients in EPIC and apply statistical ML methods, such as Bayes classifiers, boosted decision trees, logistic regression, and Generalized Additive Models (GAMs) – all of which provide interpretable and explainable results.
Specific Aim 3. Using available MRI images, we will train specially designed neural networks to classify the patients based on the historically available labeling; see Specific Aim 1. This is a compute- and data-intensive task, which will also include optimizing the architecture of the deep learning network. The performance of the neural network classifier will serve as a baseline for comparison to the following statistically explainable AI methods.
Specific Aim 4. Combining the statistical methods with the optimal features extracted by the deep learning network from the MRI images, we are aiming to develop a novel tool for the classification by utilizing a probabilistic outcome of the model.
