Most AI programs are built on “narrow” training algorithms which are pre-programmed to follow a set of rules and master a particular task, but they cannot adapt or learn from their experience. Although in many areas of AI, there are efforts to create AI that are able to build upon past experiences, there really is no replacement for human reasoning. Within healthcare if AI is to be used it has to ensure it works within a part of the team, alongside trained minds, and ultimately ‘do no harm’. Robert Wachter, M.D. in his book, The Digital Doctor: Hope, Hype and Harm at the Dawn of Medicine’s Computer Age, describes many instances of good AI systems used in healthcare. One example of this is in the field of preventing medical errors; programs like Isabel are able to minimise misdiagnosis with increasing accuracy. However, he mentions that there are also a number of poorly designed systems which have led to incorrect prescribing of medications, many of which weren’t picked up by humans until after the medication was administered, and then had to be reversed urgently to prevent physical harm.
Isabel Healthcare has been working within the field of healthcare decision support for the past 15 years and has developed a successful clinical decision support system and patient symptom checker based on natural language algorithms, with 96% accuracy of finding the right diagnosis. Isabel has all the concepts required to be considered as a step towards ‘Artificial General Intelligence (AGI)’ which is a broader form of AI. If AI is the ability to perform set tasks in a specific order, AGI is the ability to know what tasks need doing, which of those tasks is the most important, and perform them in that order.
AGI software has to be able to interact within the field it has been designed for, in this case a medical environment, and also detect and respond to a hazard, and has these core functionalities at the heart of its design:
Isabel’s system works off a evolving medical knowledge base which interacts with a natural language program. The program has been adapted to capture the essence of how humans understand this medical knowledge when they read textbooks or medical articles. This algorithm has also been further adapted, depending on whether the user is medically trained and is using the professional Isabel version, or whether they are using the Isabel Symptom Checker as patients. Isabel does not use a rules based question and answer format which is utilised by most decision support programs and symptom checkers, as it can reason by returning a list of potential diseases based on the variety of symptoms entered even if all symptoms are not present in all diseases. Rules based systems, for example, have to rely on all symptoms being present to return a potential list of causes. Isabel’s system will look for patterns within the symptoms entered but not all of them have to be present in the differential diagnosis list returned to bring it up as a diagnosis. Likewise, there could be a “red herring” symptom amongst those entered, or it could be that the symptoms entered are caused by more than one disease, and so a differential will encompass these multiple diseases.
Isabel constantly learns by new data it receives in the form of knowledge created in its database which is added by humans. As it is mimicking the human thought process, humans have to be involved to evaluate its input and output and ensure all the Isabel algorithms are functioning as they should be with the knowledge Isabel is being sent.
All the various software components work together and can be easily integrated into electronic health records and patient portals to give ease of access to Isabel’s differential diagnosis list. This can be used for further discussions with healthcare providers as a patient, or could be used as a checklist to be produced during the patient/physician encounter so that as details of the physical exam and history are gathered they can be added to and can refine that list. Isabel’s don’t miss and red flagged diagnoses are also clearly indicated to ensure they are noted during the workup or to help guide the patient as to where they should seek treatment via the Where to now? triage function.
The field of Artificial General Intelligence within healthcare is extremely interesting and will grow exponentially over the next few years, as long as systems developed work symbiotically with health care professionals, and are developed to work as the human brain does, not instead of. We cannot do away with humans, they are needed to hone this technology and take it to its full potential.
Have you read our whitepaper on the integration of systems like Isabel into medical institutions?