When caring for a critically ill child, two simultaneous thoughts are competing – the urgent need for a diagnosis to optimize treatment and the need for thoroughness – to carefully review all the possibilities. Don’t jump to a conclusion but don’t get lost in the weeds keeping the patient, and the others behind them, in limbo. We commonly see accuracy and speed as a dichotomy. This has certainly been true in the past in genomics – how many variants to review? Review variants from less likely parts of the genome? Use a more restrictive filtering rule?
We had been operating in a world where deciding to use some of the heuristic shortcuts or to time limit review meant settling for less than optimal accuracy. Time-saving techniques left some diagnoses on the cutting room floor. These simple Pareto prioritizations that are highly effective in dealing with everyday clinical situations are inherently problematic in the rare disease world. We cannot eliminate the zebras when we know it’s unlikely to be a horse.
But what if instead of speed we focus directly and exclusively on accuracy — on the most exhaustive analysis to find the best right answer without regard for time or resource. What if we take no shortcuts and consider every possibility against all known data sources using multiple algorithms. And then tune this engine for ever-higher accuracy, testing it against and learning from the best minds and proven answers. A human could never consider doing this but it turns out that in a world of Moore’s law continuously improving our cloud computing economics, this no-holds-barred approach to resources is possible and actually practical. More than practical, diagnoses are made that couldn’t be achieved by a human without the AI – there are too many orthogonal types of data at too high a quantity. The prior probability set is simply too vast.
For the patients being cared for by Fabric Genomics customers, this approach is giving diagnostic yields better than previously imagined. In about 90% of the cases, the causative variant is ranked first, in 97% of cases, that variant is in the top 5 and in virtually all cases in the top 10. The software makes the data sources and logic fully transparent to the clinician allowing for high diagnostic confidence without the need to review pages of results. The equation is turned upside down – the software exhaustively reviews everything from the bottom up so the doctor can go quickly from the top down. This change in practice is analogous to people’s initial experience with google search results – it is rare to go beyond the top few hits – it’s fast to use because it’s accurate. Within several weeks of installation, GEM users are typically reviewing under 5 variants for the vast majority of their cases. We are now achieving accuracy that gives us the speed we urgently need clinically for our sickest patients and which can be generalized to enable the future of genomics across the population.