Output of AI Algorithms
INPUTS —> MODEL —> OUTPUTS
Output
When we think about the output of the model, it all depends on how the model was designed to provide information. It can do this in many ways. If a model is a classifier, it might be designed to classify multiple elements or only one element. For example, if a binary model were designed to determine if an image contained a dog, the model will only provide one of two outputs — “yes” or “no.” If a model is a multi-class classifier, designed to recognize multiple types of animals, its output will be the name of one of the animals from the list of animals it was trained on. Some models also provide a probability score for their outputs. Other models, such as the reinforcement learning models discussed earlier, might provide a recommended action step. Below are three common categories of model outputs:
Classification
Often used in diagnostic applications.
Examples: Classifying a chest X-ray as abnormal, or a CT scan as indicative of the presence of a tumor, or a retinal fundus photograph as negative for diabetic retinopathy
Risk Prediction
Often used for risk stratification, to identify individuals who might benefit from intervention.
Examples: Predicting the risk of experiencing an adverse drug event, of being readmitted to the hospital within 30 days, or the risk of mortality from a cardiovascular event
Recommendation
Based on a desired outcome, this type of model provides a recommended intervention, action, or series of actions.
Examples: For a given patient, providing the recommended dose of IV heparin most likely to achieve target serum levels, or the initial drug combination that will be most effective at reducing symptoms of depression
And finally…. Model Deployment
Model deployment is all about implementation in practice. How would the model be operationalized in a complex clinical workflow? There are many factors that must be considered, including regulation, ease-of-use, trustworthiness, and understandability by the end-user. The first question that must be answered, however, is what role the AI model will play. There are two main approaches when implementing an AI model:
AI Model Replaces a Human Process
An autonomous system is a stand-alone system that does not require human input. It often replaces a repetitive, well-defined, and/or standardized task.
Example: Closed-loop insulin pump that automatically adjusts insulin based on continuous glucose monitoring.
AI Model Augments a Human Process
Often referred to as "human-in-the-loop" models, these models generate information to be used by a clinician in their decision-making process. The clinician acts as an important checkpoint to validate the model output and its applicability to the individual situation.
Example: Clinical decision support tools
It is worth noting that for the purposes of our blogposts, we are mostly focusing on AI applications created to be used within healthcare, whether from an operational, administrative, clinical practice or clinical research perspective. Model deployment is a critical aspect of these types of AI applications. However, another area of AI/ML that applies to health is knowledge discovery, which is an important field of research. This is an area of analytics research in which machine learning is applied to big data to generate new insights that can then inform the subject of future research and hypotheses. For example, by using machine learning to analyze >60,000 prescriptions of more than 200 drugs, several potential drug-drug interactions were discovered that had previously been unknown.¹ When used solely for knowledge discovery, deployment is not a major consideration for these AI models.
References:
Hansen Peter Wæde, Clemmensen Line, Sehested Thomas S.G., et al. Identifying Drug–Drug Interactions by Data Mining. Circ Cardiovasc Qual Outcomes. 2016;9(6):621-628.