The use of AI in healthcare is gaining increased attention with the significant advances and widening clinical use in radiology and pathology and now increasingly in genomics. In all of these cases, there are vast quantities of data to consider that should, in fact, be considered as they could be clinically significant.
Beyond the technology in use, we do see certain diagnostic situations that have a need for AI interpretation assistance. To illustrate this, we provide two different diagnostic scenarios. For instance, in adult critical care the doctor’s diagnostic process supplemented by well-researched rubrics has proven resilient. These cases are rarely primarily genetic in nature and are much more impacted by natural aging, environment, infectious agents, and lifestyle. There is a fairly high concentration of causation and a Pareto-like distribution. In contrast, the NICU pattern is quite different, with much more likelihood of a direct genetic cause. NICU genetic conditions are frequently rare and require unique considerations.
Let’s take a prototypical case – an elderly man presents in the ER with shortness of breath. In this case, there are a few highly likely and perhaps 200 possible causes and the doctor has them roughly mentally ranked in order of frequency. The clinician reranks likely causes real-time based on information as it is revealed – history, demographics, test results, etc. The top few causes make up over 95% of the cases and can be selected with reasonable confidence and confirmed via additional testing. Perhaps a few additional low-probability but high-risk causes are tested for, and a diagnosis will be confirmed in the vast majority of cases. A large but manageable dataset is analyzed and iteratively reanalyzed by the doctor in what is essentially a human Bayesian process – adjusting the prior probability based on real-time data. How likely is it heart failure (as opposed to an infection) given that there is no fever? How does the likelihood change given a specific test result? And for the vast majority of diagnoses, the doctor has confirmed and managed that specific diagnosis many times before in their career.
By contrast, let us consider the case of a critically ill child in the NICU. While there are a few common causative elements such as preterm birth, in over 30% of the cases the child’s condition has a genetic basis. The fact that there is a high chance of a genetic cause immediately brings us into a different diagnostic equation implying many thousands of potential causes. The vast majority of genetic causes will not be identified by standard newborn assay screening as they are not on the standard diagnostic panels (and those take weeks). Even the most experienced pediatricians will have only seen and be personally familiar with a tiny minority of those diseases and, on the off chance that they are familiar with a particular genetic disease, the phenotypic presentations are often not fully expressed in the newborn.
The recent Baby Bear study1 led by Rady Children’s Hospital clearly showed the prevalence of rare genetic diseases in NICU cases. Per Appendix A of the study, “Thirty-five of the diagnosed genetic diseases are rare conditions with an incidence of less than one in one million births. Sixty-five of the 71 primary genetic diseases were diagnosed just once in the Baby Bear population.” Of course, given the rarity of these conditions, it is beyond the likely human clinical experience.
We also know that the choice of treatment matters greatly. Genetic, metabolic, and neurological disorders are highly specific, and the wrong or delayed intervention can have life-long consequences.
In the NICU, we ideally would aggressively seek all reasonably accessible diagnostic information and immediately explore all possible genetic causes rather than work our way slowly along a curve that lacks the steep Pareto shape. Time is often the enemy in these cases. Damage from seizures, nutrition acting as a metabolic poison, or invasive procedures can take place that could be avoided with an early, specific diagnosis.
Just as we treat a critically ill febrile patient with broad-spectrum antibiotics to not lose time with narrow therapeutic shots in the dark, we need a broad but fast approach to diagnosis.
Fortunately, with technology lowering the cost of NGS and with the support of AI algorithms such as Fabric GEM, this approach is coming into use. Multiple peer-reviewed studies2 including Farnaes et al 2018 show significant clinical efficacy of this approach and even cost savings.3 The technology is here, it’s available and even economical. Now is the time to insist on its use for the NICU babies that depend on us.