Despite enormous advances in diagnostics, genetics, and medical devices, 5-15% of patients are still misdiagnosed. To solve this problem, we must minimize intuitive decisions and “thinking fast” in medicine.
White spots of medical progress
If we don't start using artificial intelligence in healthcare, we won't be able to reduce the number of medical errors, which aren't decreasing despite rising healthcare expenditures – according to Gaurav Singal and Anupam B. Jena from Harvard Medical School in their opinion published recently in the Los Angeles Times. The call for implementing AI in healthcare follows another article by American scientist and author Eric Topol in The Science (“Toward the Eradication of Medical Diagnostic Errors”).
The numbers are alarming. In the U.S., up to 15-25% of diagnoses are incorrect; in Europe, the percentage is estimated at 5-15%, but the actual numbers might be even higher. Misdiagnoses lead to drug prescribing errors, delayed treatment of patients, deaths, avoidable complications, and costs. A 2015 National Academies of Sciences, Engineering, and Medicine report found that diagnostic errors affect 5% of the population annually. Everyone is likely to experience such an error, often unknowingly, in their lifetime.
Gut feelings have no place in medicine
Where is the problem? The honest answer lies in diagnostic procedures and workflows that have remained unchanged for the past century. Advanced medical imaging methods, genetic testing, and digitization allow doctors to access precise data on a person's health status following 360-degree patient view, data-driven medicine, or integrated medicine principles.
The availability of medical knowledge has steadily improved over the past years. Since data is collected in electronic medical records (EHRs), doctors can open a patient's complete medical history with a single click, regardless of where they work and where the patient was treated before. Despite these significant advances, one thing hasn't changed in hundreds of years: how medical decisions are made. The bad news is they are made manually. According to Gaurav Singal and Anupam B. Jena, “diagnosis has largely remained a human endeavor, with doctors relying on so-called illness scripts—clusters of signs, symptoms, and diagnostic findings that are hallmarks of a disease.”
There is nothing wrong with using intuition when diagnosing widespread diseases like flu or skin rash. But the more complex the disease is, the bigger the probability of making a mistake. The human brain has cognitive limitations; thus, having access to more data does not automatically mean better decisions. Doctors are also susceptible to many biases – including confirmation bias, overconfidence bias, anchoring bias, and availability heuristics – because they rely on intuition instead of facts. This makes it easy to overlook or ignore information crucial to a correct diagnosis, no matter how complete and accessible the data in the EHR is.
The problem is not only our cognition but also the healthcare system itself. Doctors cannot analyze all the data because they work under time pressure due to the growing demand for medical services and existing business models. The rush is the primary source of mistakes, causing doctors to rely on quick, intuitive thinking (System 1, according to psychologist and Nobel-Prize winner Daniel Kahneman) and neglect analytical thinking based on data and medical literature (System 2). “With the brief duration of a clinic visit, it is not surprising that there is little time to reflect because it relies on System 1 thinking, which is automatic, near-instantaneous, reflexive, and intuitive,” claims Eric Topol.
AI will turn on System 2 and switch off System 1
However, artificial intelligence can change it by bridging the gap between the amount of data collected and its actual use in the diagnosis process. AI doesn't get tired, can analyze gigabytes of information in a fraction of a second, and is less susceptible to factors leading to medical errors. It can pick up hidden patterns in data, pixels of change on medical images, and small but important records in medical files from several years ago that the human eye cannot detect.
So far, scientific studies confirm that AI identifies strokes in seconds after the examination, much faster than radiologists. It can predict the onset of sepsis in hospitalized patients based on hundreds of vital signs and medical history and correctly diagnose pneumonia or kidney damage. Multimodal AI analyzes data such as texts and images in a patient's EHR, helping to identify rare diseases. The doctor makes the final decision, but it is then based on facts, not solely on System 1.
Technology is there, but it still must be refined
There is still a shortage of high-quality data to train accurate algorithms. Without them, AI can be biased and make mistakes according to the “garbage in, garbage out” principle (lousy quality of data in, low diagnosis quality out). Another obstacle is cost – currently, only large university centers can afford to implement AI. Machine learning algorithms are still expensive, and in most cases, their use is not reimbursed by payers. On top of this, many medical facilities are discouraged by complicated regulations regarding the use of AI algorithms and the use of data to train them.
AI-assisted diagnostics are seen as the greatest hope in the fight against medical errors. Used as a source of second opinions and physician support, it will make any diagnosis based on objective data. For doctors, this means more secure and evidence-based decision-making. For patients, it means safer healthcare.
Daniel Kahneman once said he would never trust a doctor who relies only on his intuition. For millions of years, the dominance of System 1 over System 2 has allowed us to survive as a species. It’s better not to think too long if the rustling leaves suggest a tiger that wants to eat us. However, medicine is a fact-based science that requires an objective evaluation of facts and data, which doctors, unfortunately, don't have the time for, but AI fortunately does.
„We are certainly not there yet. But in the years ahead, as we fulfill the aspiration and potential for building more capable and medically dedicated AI models, it will become increasingly likely that AI will play an invaluable role in providing second opinions,” concludes Topol. Gaurav Singal and Anupam B. Jena agree: "This could be a special moment for diagnosis if we invest enough and do it right.”
