Image from Adobe
Healthcare data is often incomplete and inconsistent, limiting efforts to improve patient outcomes and operational efficiency. A 2021 report from Sage Growth found that only 20% of healthcare organizations fully trust their data. Because records follow patients across providers with shifting identifiers and coding schemes, the same encounter often appears multiple times or partially, breaking longitudinal linkage and skewing analyses.
All of this complicates developing drugs to tackle unmet medical needs. “Disconnected and inconsistent data really limits the evidence generation that is relevant for drug development,” said Glynn Dennis, Chief Science Officer of Kythera Labs, which provides a data management and analytics platform as well as robust datasets to pharma and biotech companies.
Remastering healthcare data
Glynn Dennis
A large part of the inconsistency and duplication of healthcare data comes from the patient journey, Dennis explained. When patients move or switch providers, or change their name if they get married, their patient data can become disconnected and disorganized. That is where Kythera’s services start when remastering data.
“It’s really all about building that contiguous patient journey with high resolution and not just high quality, but an understanding of where quality is good and where it is bad,” Dennis said.
Kythera’s Wayfinder platform can take datasets that have tens of petabytes of raw data and turn them into what Kythera calls “patient events assets,” which are episodes of patient care, Dennis said. These episodes can include operations and other treatments, placed within the geographical context of providers.
From these episodes, Wayfinder derives assets such as chronic condition assets or oncology assets. These assets are “fit for purpose” data that has gone through the full remastering process and are “analytics and AI ready,” Dennis said.
Kythera provides various datasets for client use, but the company also works with biotech and pharmaceutical companies that have their own clinical and real-world data. In these cases, Kythera’s technology can be deployed into their data environment to perform tokenization and remastering as well as harmonizing their data with Kythera’s to create a complete, remastered dataset that is ready for analysis.
Data is moving upstream
Kythera’s services are used for observational evidence generation for safety studies and health outcomes research. These use cases “support outcomes of different drugs, have published the observation of comorbidities associated with a primary condition based on treatment regimens and look at safety profiles,” Dennis said.
But observational studies are not the only way Kythera’s services are being used. Increasingly, pharmaceutical companies are connecting patient sample multiomics data to real-world data to gain a deeper understanding of the medical history of the patients, instead of collecting real-world data only after a drug has been approved to observe its effects.
“The trend right now is that real-world data is moving further and further upstream in the clinical development process, even into the preclinical space,” Dennis said.
As sponsors push real-world data earlier into development, what matters is task-specific input quality: are encounters linked to one patient, codes/units harmonized, and labels faithful to the record? Remastering improves patient-journey linkage and label fidelity, which in turn lifts cohort selection, outcomes modeling, and any retrieval-augmented analysis over clinical corpora.
In addition to the analysis of data, AI is increasingly able to perform higher-level tasks, although critics and insiders such as Google DeepMind’s Demis Hassabis note that large language models can still be brittle on both complex and simple tasks. In August, Hassabis said, “It shouldn’t be that easy for the average person to just find a trivial flaw in the system.”
In parallel, the emergence of ‘agentic’ AI systems are beginning to have an influence in fields ranging from software engineering to drug development. In pharma, for instance, agents’ ability to collaborate shows up in multi-agent frameworks that autonomously cycle through design-make-test-analyse loops. A recent preprint reported a tripling of hit-finding performance over baseline in drug-design tasks. “I think there’s a real opportunity for AI to reduce the time to move the pipeline forward,” Dennis said, “It’s about all of the stuff in between the science, which is all the operational things that are right now very human-oriented and manual. Anything that’s very repeatable, such as documentation, clinical study protocols and transformation of data, is ripe for disruption.”
