ABSTRACT
The application of artificial intelligence (AI) to healthcare has garnered significant enthusiasm in recent years. Despite the adoption of new analytic approaches, medical education on AI is lacking. We aim to create a usable AI primer for medical education. We discuss how to generate a clinical question involving AI, what data are suitable for AI research, how to prepare a dataset for training and how to determine if the output has clinical utility. To illustrate this process, we focused on an example of how medical imaging is employed in designing a machine learning model. Our proposed medical education curriculum addresses AI's potential and limitations for enhancing clinicians' skills in research, applied statistics and care delivery.
The application of artificial intelligence (AI) to healthcare has generated increasing interest in recent years; however, medical education on AI is lacking. With this primer, we provide an overview on how to understand AI, gain exposure to machine learning (ML) and how to develop research questions utilizing ML. Using an example of a ML application in imaging, we provide a practical approach to understanding and executing a ML analysis. Our proposed medical education curriculum provides a framework for healthcare education which we hope will propel healthcare institutions to implement ML laboratories and training environments and improve access to this transformative paradigm.
Subject(s)
Artificial Intelligence , Education, Medical , Delivery of Health Care , Humans , Machine LearningABSTRACT
BACKGROUND: Coronary CT angiography (CCTA) and contrast-enhanced thoracic CT (CECT) are distinctly different diagnostic procedures that involve intravenous contrast-enhanced CT of the chest. The technical component of these procedures is reimbursed at the same rate by the Centers for Medicare and Medicaid Services (CMS). This study tests the hypothesis that the direct costs of performing these exams are significantly different. METHODS: Direct costs for both procedures were measured using a time-driven activity-based costing (TDABC) model. The exams were segmented into four phases: preparation, scanning, post-scan monitoring, and image processing. Room occupancy and direct labor times were collected for scans of 54 patients (28 CCTA and 26 CECT studies), in seven medical facilities within the USA and used to impute labor and equipment cost. Contrast material costs were measured directly. Cost differences between the exams were analyzed for significance and variability. RESULTS: Mean CCTA duration was 3.2 times longer than CECT (121 and 37 âmin, respectively. p â< â0.01). Mean CCTA direct costs were 3.4 times those of CECT ($189.52 and $55.28, respectively, p â< â0.01). Both labor and capital equipment costs for CCTA were significantly more expensive (6.5 and 1.8-fold greater, respectively, p â< â0.001). Segmented by procedural phase, CCTA was both longer and more expensive for each (p â< â0.01). Mean direct costs for CCTA exceeded the standard CMS technical reimbursement of $182.25 without accounting for indirect or overhead costs. CONCLUSION: The direct cost of performing CCTA is significantly higher than CECT, and thus reimbursement schedules that treat these procedures similarly undervalue the resources required to perform CCTA and possibly decrease access to the procedure.
Subject(s)
Computed Tomography Angiography , Medicare , Aged , Coronary Angiography , Humans , Predictive Value of Tests , Tomography, X-Ray Computed , United StatesABSTRACT
BACKGROUND: Pericoronary adipose tissue (PCAT) attenuation has been identified as a marker for cardiovascular risk. The effect of contrast enhancement on fat attenuation is unknown. We aim to compare precontrast coronary scans to postcontrast CCTA for quantification of pericoronary fat volume and attenuation. METHODS: Thin slice pre- and post-contrast studies obtained at 120â¯kVp, heart rate <60, with no plaque or artifact in the right coronary artery (RCA) were selected. Analysis was limited to pixels -30 Hounsfield units (HU) to -190 HU and from 10â¯mm to 50â¯mm distal to the RCA origin at a radial distance equal to the vessel diameter. A subgroup with no plaque across all coronaries was also analyzed. RESULTS: Of 119 study pairs, the average RCA diameter was highly correlated at 3.85â¯mm (postcontrast) and 3.84â¯mm (precontrast), râ¯=â¯0.97, pâ¯<â¯0.0001. The mean attenuation of pre- and postcontrast images was also highly correlated at -87.02⯱â¯7.15 HU and -82.74⯱â¯6.54 HU, respectively (râ¯=â¯0.65, pâ¯<â¯0.0001). Pericoronary fat volume in the -190 to -30 HU range was 396â¯mm³ lower in the post contrast versus pre-contrast, consistent with higher attenuation (less negative) voxels postcontrast (pâ¯<â¯0.0001). Inter- and intra-reader agreement ranged 95-100% and 90% for precontrast and 85-90% for postcontrast studies, respectively. Subgroup analysis revealed precontrast attenuation -85.59⯱â¯7.53 HU and postcontrast -82.21⯱â¯7.15 HU were highly correlated râ¯=â¯0.67, pâ¯<â¯0.0001. CONCLUSION: Pericoronary fat enhances with iodinated contrast, potentially explaining some of its risk-predictive capabilities. Fat attenuation and volume can be reliably measured from precontrast calcium scans, with volume quantification showing particularly strong correlation. Excellent inter- and intra-reader agreement is also demonstrated.
