EXPLAINABLE AI IN HEALTHCARE

Explainable AI (XAI) refers to artificial intelligence systems that provides an explanation of the decisions of the system. Medical decisions are life changing and carry significant implications. It is therefore important for both the patient and doctors to know how the system has arrived at a decision such as a diagnoses or treatment recommendation. Yet (AI) has great benefits. The system can help make decisions fasters and can analyze huge amounts of data that is beyond human ability. An AI system can easily predict atrial fibrillation from an electrocardiogram.

Explanations are also important to foster trust on the system as well as between patients and doctors. An explanation can be used to improve the system model and remove bias. AI systems are built upon training data may not be representative of the presented case leading to a biased decision. A good example is a hospital system for detecting skin cancer that has been trained mainly on Caucasian skin data rather than dark-skinned people. Such a system could easily miss to accurately detect cancer on dark skinned people leading to serious consequences. If the diagnosis comes with an explanation then in could reveal this biased training data and help make the necessary changes.

There are three broad categories used in explainable AI techniques. They are feature importance, counterfactual explanations, and natural language explanations

Feature importance: A hospital admission systems can use certain features such as age, past dismissions etc to recommend a patient readmission. In the case of the skin cancer detection system, it may use features like the size, shape and color of the lesion to make a diagnosis. The explanation should then show why the diagnosis is made and the doctor would have an opportunity to explain or review.

Counterfactual explanations: This is similar to "what if” scenario analysis by showing potential impact of changing some parameters such as age, dosage, gender etc. It would for example show that a person in their 80s would respond differently. 

Natural language explanations: Here the emphasis is to provide plain language explanations that patients and doctors understand. It can come in handy with medical devices such as a heart pacer changing the heart rate because the patient heart late has become too slow.

Explanations should be transparent and meaningful. A patient may only need to know that their the pace maker increased the heart rate that was dropping but the cardiologist can get a more detailed explanation.  The explanation should also be accurate and consistent based on the model data otherwise it will become confusing and undermine trust. It should be able to tell when the wrong parameters are used for its model. 

Explainable AI is a very useful and growing field. It has great potential in healthcare.