In Tom Cruise’s film Minority Report, which was adapted from the eponymous science fiction novella, “Precrime” is a predictive policing system dedicated to apprehending and detaining people before they have the opportunity to commit a crime. "Punishment was never much of a deterrent and could scarcely have afforded comfort to a victim already dead" goes the protagonist. At their best, very early warnings of impending disease can function in much the same way. When we detect upstream warning signs of a disease, we may have the ability to either mitigate the associated risk, arrest the progression of the disease, or slow the disease down before it manifests in sickness. Indeed, early detection of disease has been a goal of physicians for as long as medicine has been practiced; it is also a goal that AI increasingly helps them achieve.

When diagnosing disease, time is everything. Detecting the presence of risk early gives clinicians and patients alike the time they need to design and implement an effective strategy for managing downstream risk and for treating the disease. AI and deep learning hold enormous promise when it comes to detecting early signals of disease. This is due in part to their ability to quickly comb through large volumes of multi-modal patient data, but it is also due to their ability to detect non-obvious, non-linear correlations between patient features and disease, thereby uncovering previously unacknowledged warning signs. However, the correlations and indicators of disease detected by deep learning algorithms are not foolproof. Rather, they serve to alert clinicians to the possibility of disease, which can then be examined or ruled out by a physician. The risk-sensing mechanisms of the future will enable exactly this kind of interaction between patient, clinician, and artificial intelligence to support the early detection, investigation, and treatment of the risk of disease.

Pre-term birth is a significant driver of disease burden for both mothers and children, both globally and within the United States. Babies born prior to 37 weeks in utero require more expensive care, and more of it, than babies born after 37 weeks. CareCentra’s Healthy Maternity program uses a mobile pregnancy tracking and coaching app to reduce the risk of pre-term birth in diverse cohorts of mothers and babies. The goals of the program are a) to help those with a known risk of pre-term birth to stay pregnant until the 37^th week of gestation, and b) to detect the risk of possible pre-term births among participants with unknown risk (less than 30% of risk of PTB is currently detected upfront using family history and a few other features). While outcomes for patients on our program are significantly better than the average outcomes in the geographies where we operate, the data we collect through our platform nevertheless enables robust statistical inference, leading to further improvement of outcomes. We find that patients for whom our algorithm raises multiple alerts tend to be more likely to have shorter pregnancies than those for whom it raises fewer alerts. An alert can be triggered by several different factors, including the failure to report blood pressure reading, repeated non-adherence to prescribed meds if any, or patient-reported adverse symptoms. This shows that by minimizing the number of trigger events that raise alerts (i.e. acting preventatively to avoid alerts), patients and care teams can work together to reduce the risk of pre-term birth. At CareCentra, we’re able to replicate this same mechanism in several different clinical contexts, from diabetes to respiratory disease.

We performed a LASSO (least absolute shrinkage and selection operator) regression and feature selection procedure on over a dozen potential predictors of the gestational age of a maternity patient’s child at birth. This technique learns a linear model relating independent to dependent variables, while enforcing the requirement that the sum of the magnitudes of all coefficients be no greater than a regularization parameter. This requirement forces the regression coefficients for unimportant predictors to zero, leaving only the important predictors of the dependent variable – in this case, the length of a patient’s pregnancy. After using cross-validation to find an optimal regularization parameter, we found that the number of alerts raised in our app was one of only a small number of features that predicted pregnancy length, with each alert associated with a 0.46 day decrease in the gestational age of a patient’s child at birth. This should be seen in juxtaposition to the cost of PTBs – a late PTB in the 35^th or 36^th week of gestation costs the system $50k and an early PTB that is 100 days too early costs about $1M. Thus, every day a high-risk mom stays pregnant after 24 weeks saves the system $10 k. This suggests that caregivers seeking to extend patient pregnancies to their full term should aim to prevent the types of events that lead to alerts, while also responding diligently to those alerts that are generated.

AI-supported early-warning systems hold great promise for treating patients with a wide variety of conditions. However, not all such systems are created equally, and one must be diligent in ensuring that the alerts generated by an application are valid measurements of risk. At CareCentra, we take this diligence seriously, bringing statistical rigor to the question of whether the alerts raised by our app are credible indicators of potential disease. The evidence to date suggests that they are. This kind of analysis is crucial to establishing a mutual trust between caregiver, patient, and technology that is necessary to leverage the power of artificial intelligence and deep learning in a healthcare setting.