Flahy’s bet is that the work sitting between a patient’s record and the next treatment decision belongs on a graph, not in a pile of retrieved text.
Jagjit Singh, the company’s chief executive, made that case to SiliconANGLE’s theCUBE. In his telling the knowledge layer carries the load: it is what wires raw data to a model, and what tells that model which facts matter for this individual and what any single data point implies.
A graph-based information database, plus a system trained to spot relationships among data points, are what years of work produced, Singh said. Clinical facts, under this design, are not text to be summarized.
He offered a hypothetical to show why. Take someone carrying a particular genetic mutation and a high cholesterol reading. Along comes a new clinical finding, and the obvious question is whether it should shift that person’s care. Answering it, he argued, is a traversal problem rather than a language one.
Neo4j is among the graph vendors Flahy works with, and the clinical reasoning itself runs on engines the company built.
Wearables remain an open problem. The hard engineering lies in folding a stream of continuous device readings into the same graph as clinical records, so the system can read change over time.
That architecture also explains the avoidance of plain retrieval: two facts that interact can be expressed in a graph, while a stack of retrieved passages cannot say the same.
