Patients using AI tools on their own health
Patients use AI chat assistants to make sense of their own records and symptoms, and the tools hold potential at every phase from first symptoms to living with a disease. Generative tools also open new forms of patient education, while explainability, privacy, accuracy and regulation bear on how far patients can accept and rely on what the tools tell them.
Patients' use of the tools
AI is in use at the consumer level as well as among researchers and companies: patients use it too, and everyone is a patient. In May 2024 one patient guessed that almost everyone in a conference audience had searched Google for symptoms related to their disease, and saw that far fewer had used a chat assistant for it. In his view such use is really coming, with a lot of potential and some challenges.
Needs along the patient journey
The patient journey runs from living healthy through developing symptoms, diagnosis, starting and perhaps many times changing treatment, to living with the disease, and a patient's needs differ in each phase. With symptoms, the need is to detect them as early as possible. At diagnosis, the physician gives a prescription that may or may not work, perhaps over several iterations, and the knowledge of all this sits in the head of the physician the patient relies on. With the disease, the need is to live a good life within the condition. In the same patient's view, AI has potential in every one of these phases. A patient newly diagnosed with a disease that has a complex name wants to know what it is in human language, and AI could help explain the terms, what the disease involves, how it will affect their life and its treatments.
In the same patient's view, the opportunities for patients include clarification, in the patient's own language when much of the information is in English, and personalisation to the patient's own context. For example, a patient can describe their working and family environment and physical performance to a chatbot and ask for suggestions to live a healthier life and improve their quality of life. A lot of information already comes out of such questions. A further prospect is a digital twin that continuously captures a person's data sources and, based on certain assumptions, predicts the effect of changing an input. In the same patient's view, such twins were getting closer in May 2024, and the more the omic space can be linked with behavioural science and people's data traces and models validated, the more opportunity they hold to help patients.
Records and symptoms put through a chat assistant
The same patient put his own medical record through a chat assistant. He asked his general practitioner to export the record in the format Belgian practices use, a patient migration file, a format that may differ in every country. The export was unusable to him, so he put it into ChatGPT and asked what relevant information it contained. The first answer reasoned from the file's structure and flagged sleep apnoea with its ICD code, which gave a little explanation but was still not very understandable to the average person. Asked what that meant in simple terms, the tool gave a much clearer explanation: sleep apnoea is a condition in which people stop breathing for short periods while asleep. He stripped his own and his physicians' personal data before uploading the file, and still did not know where it had gone. In his view, many patients are already using these opportunities.
He also used the tool on a symptom. During a corticosteroid dose-reduction trial he developed a chronic cough that lasted about five months, stopped him speaking in meetings and ruined his sleep, and that repeated consultations with the physician had not resolved. He had thought it most probably came from the dose reduction, the corticosteroids having suppressed the cough until then. He described to ChatGPT what he did, what he took and his history, and asked for suggestions to get rid of the cough, and the cause turned out to be an ACE inhibitor, a known cause of chronic cough. Thanks to the tool he took the matter to the trial's principal investigator, the medication was changed, and the cough disappeared after four weeks. A patient can use the technology in this simple way.
Patient education
In the same patient's view, generative tools also open new forms of patient education. One is a deepfake avatar video in which a patient's own physician, with her face and voice, could deliver education targeted to that patient's context. A physician with too little time to explain everything could hand it over as a QR code that the patient scans at home to play the videos on certain topics. Such videos are already being generated, and the same patient had made some himself. A step further is a dynamic avatar that a patient can put questions to. Such technologies also translate, so a physician in Brussels could talk to a patient who came from Ukraine without speaking their language, which is an opportunity to increase equity, diversity and inclusion.
Acceptance of what the tools say
A research project for people with COPD, a population that is statistically very old, was being set up in May 2024 on how the use of AI models affects their acceptance of the information that comes out and whether they need to understand it. In its scenario a patient is told that, based on their data and the models, they will have an exacerbation in two weeks and is offered options, and the question is whether they may trust the message more if it comes through a physician, or less if the decision is in their own hands. In the same patient's view, explainability therefore becomes very important.
Risks
In the same patient's view, the opportunities for patients come with potential negative effects to be mindful of. Data privacy is one. Accuracy and reliability are another. Evidence increasingly shows that the tools perform quite well, but nobody is ever a hundred per cent sure, and a medical decision taken on bad information can have bad consequences, which raises the question of how patients are prepared for that. Regulation is a third. The medical device regulations, the GDPR and HIPAA raise the question of how far such tools can go and what validation they need before they reach patients. That includes whether a general chatbot such as ChatGPT, which carries some disclaimers but returns medical suggestions, should count as a medical device, and whether patients should be told to stay mindful of the risk whenever they use one.
AI tools may also need an equivalent of pharmacovigilance, which would capture their adverse effects so that they can be learned from collectively. In May 2024 the same patient remained very positive: in his view the tools will only become better for patients, the opportunities will outweigh the potential negative consequences, and society is responsible for keeping them steered in a good direction.