Automated Decision Support System for Traumatic Injuries

Abstract

With trauma being one of the leading causes of death in the U.S., automated decision support systems that can accurately detect traumatic injuries and predict their outcomes are crucial for preventing secondary injuries and guiding care management. My dissertation research incorporates machine learning and image processing techniques to extract knowledge from structured (e.g., electronic health records) and unstructured (e.g., computed tomography images) data to generate real-time, robust, quantitative trauma diagnosis and prognosis. This work addresses two challenges: 1) incorporating clinical domain knowledge into deep convolutional neural networks using classical image processing techniques and 2) using post-hoc explainers to align black box predictive machine learning models with clinical domain knowledge. Addressing these challenges is necessary for developing trustworthy clinical decision-support systems that can be generalized across the healthcare system. Motivated by this goal, we introduce an explainable and expert-guided machine learning framework to predict the outcome of traumatic brain injury. We also propose image processing approaches to automatically assess trauma from computed tomography scans.

This research comprises four projects.

In the first project, we propose an explainable hierarchical machine learning framework to predict the long-term functional outcome of traumatic brain injury using information available in electronic health records. This information includes demographic data, baseline features, radiology reports, laboratory values, injury severity scores, and medical history. To build such a framework, we peer inside the black-box machine learning models to explain their rationale for each predicted risk score. Accordingly, additional layers of statistical inference and human expert validation are added to the model, which ensures the predicted risk score’s trustworthiness. We demonstrate that imposing statistical and domain knowledge “checks and balances” not only does not adversely affect the performance of the machine learning classifier but also makes it more reliable.

In the second project, we introduce a framework for detecting and assessing the severity of brain subdural hematomas. First, the hematoma is segmented using a combination of hand-crafted and deep learning features. Next, we calculate the volume of the injured region to quantitatively assess its severity. We show that the combination of classical image processing and deep learning can outperform deep-learning-only methods to achieve improved average performance and robustness.

In the third project, we develop a framework to identify and assess liver trauma by calculating the percentage of the liver parenchyma disrupted by trauma. First, liver parenchyma and trauma masks are segmented by employing a deep learning backbone. Next, these segmented regions are refined with respect to the domain knowledge about the location and intensity distribution of liver trauma. This framework accurately estimated the severity of liver parenchyma trauma.

In the final project, we propose a kidney segmentation method for patients with blunt abdominal trauma. This model incorporates machine learning and active contour modeling to generate kidney masks on abdominal CT images. The resultant of this component can provide a region of interest for screening kidney traumas in future studies.

Together, the four projects discussed in this thesis contribute to diagnosis and prognosis of trauma across multiple body regions. They provide a quantitative assessment of traumas that is a more accurate measurement of the risk for adverse health outcomes as an alternative to current qualitative and sometimes subjective current clinical practice.