What should we do if computers predict who may develop a terminal disease, attempt suicide or contract diabetes? And what should you or your doctor do on receiving such a forecast? Is it even possible? Yes - the growing accuracy of computer algorithms suggests we will soon be facing such challenges.
Researchers at Florida State University, for example, recently mined data from electronic health records in which they identified a group of patients who had attempted to commit suicide. They then used the fast-growing data analysis technique, machine learning, to identify a combination of factors that could most accurately predict a suicide attempt.
This kind of predictive modelling is not new, but the capabilities of computers to recognise complex interactions among variables associated with outcomes of interest has greatly advanced in recent years. Computers can now not only analyse huge volumes of data, they are also capable of reading and understanding free text (such as a doctor's dictated note) and can combine multiple sources of information to achieve accurate predictions. Using machine learning, these systems are also designed to continuously improve as they gain experience with the outcomes they are measuring. As a result, they produce timely insights that exceed the capacity of humans.
The Florida State University researchers found the machine learning algorithms could predict a suicide attempt with 80 to 90 percent accuracy up to two years in advance. Other organisations are developing similar approaches for a variety of other conditions such as depression, heart failure, heart attacks, dementia, Parkinson's disease, and chronic kidney disease.
The health industry is awash with data that can feed and inform these computer models, making it likely that their accuracy will continue to improve. Are we rapidly approaching a time when computers are so confident in their predictions we should consider that we already have the disease, or what we could call a "proto-disease"?
The promise of accurate disease predictions is they potentially allow for interventions that could reduce or eliminate the likelihood the diseases will occur. Unfortunately, the ability to predict is advancing faster than our ability to prevent. Consider a visit to your doctor during which she advises you a computer programme has predicted you will develop diabetes mellitus within the next two years. She prescribes a large dose of advice on healthy living (which you may be more likely to follow given the computer's warning) but otherwise has little to offer. Indeed, studies that have evaluated prevention of diabetes with available medications have been inconclusive. You no doubt leave the office sullen, now bearing the burden of having a proto-disease.
To further complicate matters, it may not be your doctor who breaks the news to you. Data mining can also be conducted by organisations that have access to information about you, such as your health insurer. Ideally, health insurers will use the information to help keep you healthy but as noted earlier, effective interventions are not yet fully developed.
There may also be more nefarious uses for the information. A health insurer could, for example, use information about health risks in their insured population to their commercial advantage, and their customers' detriment. Knowledge is power - especially in the insurance business.
Any insurer that can accurately predict who of their insured will develop illnesses (and hence cost more) will have an undisputed advantage when it comes to payment terms.
Ethical, moral and legal questions also arise. Does your health insurer - or more importantly your doctor - have to inform you when an accurate computer algorithm predicts you will develop a disease? Should it be considered negligent if they do not?
The near term challenge is to confirm the accuracy of the predictions and to find ways to inject them into the work-flow of already over-burdened health professionals. The predictions will ideally be accompanied by evidence-based information to help guide the clinician on what actions to take. Healthcare organisations already committed to population health (in which care is organized around groups of patients such as those with diabetes) will need to find ways to make use of the information. Finally, this data revolution will perhaps lead to a renaissance in disease prevention, with a focus on proto-diseases.
Advancement toward machine learning and disease predictions is already well under way. It's just a matter of time before their predictions become a factor in determining how best to treat patients. The key challenge will be whether effective interventions can keep pace with the predictions.