Specific used cases will be discussed of FDA approved AiR™ software and the process of FDA approval and market niche of clinical decision support systems not requiring FDA clearance. The used cases will help characterize specifics of which Ai data science tool was used to generate the algorithm and the need for evaluation of the Ai statistical evaluation tool called the confusion matrix that vendors present to show statistical performance and how to analyze the strength of the confusion matrix being shown and to empower the administrator and technologist to ask for the F1 score. Quality assessment of algorithms including monitoring drift of algorithms, creating a feedback loop to the vendor to allow for algorithm retraining and improved performance and mechanism to report Ai adversarial events.
Intercept of Ai software tools meeting AMA definition for reimbursement as expert systems and expectations for reimbursement to be discussed.
Concepts of AiR™ ethics and informed consent of the patient as the use of AiR™ in their clinical care to be discussed. Pitfalls in the application of AiR™ including automation biases where in human decision making, humans have a tendency to select a pathway requiring the least cognitive effort which can result in technology and AiR™ dictating the pathway of clinical care. Heavier workloads and an increase in workloads with increasing time pressures leads to further susceptibility of bias. Another example of biases is in the actual algorithm development which can include statistical or social biases or in some cases both.
AiR™ algorithm development includes application of machine learning as well as deep learning with subcategories of supervised and unsupervised learning, support vector machines, convolutional neural networks and graph neural networks with efforts of applying AiR™ software to do multimodal fusion that incorporates clinical history, laboratory data, genomics and imaging under an umbrella of multimodal fusion. This can allow for improved predictive analytics and outcomes incorporating for the first time unstructured data which represent the form that the majority of patient data is in.