The easiest way for a brand-new, venture-backed college graduate to build a start-up is to leverage large language models (LLMs). While uptake has lagged behind other industries, health systems and integrated delivery networks (IDNs) have slowly been incorporating this technology. Vendors have hosted private dinners and invitation-only events at conferences, and adoption of solutions using generative AI and LLMs became a reality. But rarely in life is the easiest path forward the safest.
Easy Street Is a Blind Alley
The easiest way for a brand-new, venture-backed college graduate to build a start-up is to leverage large language models (LLMs). While uptake has lagged behind other industries, health systems and integrated delivery networks (IDNs) have slowly been incorporating this technology. Vendors have hosted private dinners and invitation-only events at conferences, and adoption of solutions using generative AI and LLMs became a reality. But rarely in life is the easiest path forward the safest.
Next came the data breaches and the inevitable security risks. Models began doing things nobody instructed them to do, including hacking other generative LLMs and covering their tracks to disguise their actions. Health systems quickly reacted by establishing AI committees to monitor and govern the use of LLM-based AI models. They had been burned by breaches and by vendors who couldn't deliver on boastful marketing metrics, and they had taken risks they were once afraid to take. Now health system committees are asking the questions they should have asked at the beginning:
- Where does patient data go, and who has access to it?
- Can it be breached?
- Is it being shared with vendors and used to train their models?
- Are you selling our de-identified data (patient data with names and other identifiers removed)?
- What risks does this pose to our health system?
- Can we require vendors to carry cyber liability insurance high enough to protect us if a breach occurs, and at what cost?
As a result, AI adoption has nearly stalled. Evaluation processes now take a very long time and cover risk and security and buy-versus-build decisions. Executives also lean on platforms that have already been approved for a specific purpose, so that no sponsor has to put their career on the line if the LLM they backed jumps the tracks or allows a major breach. What health systems ultimately require is Safety, Security and Risk Mitigation.
The healthcare market has done a complete turnaround on deploying AI/LLM models and applying them across the enterprise. There are extensive risk analyses and cyber liability coverages that would put any start-up out of business after a breach, which diminishes or eliminates the value of the technology in the first place. Health systems are now very sensitive to using generative AI or LLM-enabled solutions, whether from vendors or developed internally, unless the risks can be mitigated. They also know that application solutions built on legacy software technology stacks will soon need to be replaced with safe, secure, proven and trusted AI technology.
They are looking for automation and technology solutions that deliver value without the fear inherent in LLMs. Native technology, not LLM wrappers. Safe AI-enabled software applications that run on servers behind their firewalls in their own private cloud environments, managed by a trusted team and supported by a proven and trusted partner. Patient data should never leave your system, and nobody should have access to it unless you grant it.
Safe means guardrails. Oncology is a discipline built around standards, whether it's treatment, diagnosis, or accreditation. Generative AI can learn on its own and, without guardrails, does not adhere to standards. It can go rogue and make decisions nobody instructed it to make. Safe means technology with guardrails built in. It operates within defined oncology coding rules, cannot infiltrate or be shared with your other applications, and keeps protected health information (PHI) inside your firewall. It means your cyber liability coverage and limitation of liability stay exactly where they are. Generative tools in the cloud cannot say that today, and bad actors and artificial superintelligence will exploit systems in a way that even the largest cyber liability policy won't even begin to touch.
Accuracy Alone Is Not the Argument
Inspirata AI's natural language processing (NLP) Oncology Data Intelligence Platform uses deep search, oncology-specific rules and standards, and inference to interpret ambiguous terminology. This allows the platform to automate reportable-only casefinding with 99% accuracy. But no one should buy on accuracy alone, because accuracy is a moving target. Coding rules, staging manuals and reporting requirements are revised every year, so an answer that is correct under one year's standards can be wrong under the next. A solution is only as accurate as its ability to keep up with those changes, and to show its work when it does.
Generative AI models teach themselves. Even with time, however, they will not be able to significantly close the gap on reportability, because LLMs hallucinate and work on the concept of “When I was asked this question last time, this was the answer I gave (right or wrong).”
They also cannot explain why the answer was chosen, what standard or rule is behind it, or what other selection or validation criteria were used. Over weeks and months, LLMs can find a data element to put in every field of a North American Association of Central Cancer Registries (NAACCR) abstract. The Oncology Data Specialist (ODS) will still have to search electronic medical records (EMRs), process documents, and spend the same amount of time reading reports to correct or validate each field.
Inspirata AI's NLP Oncology Data Intelligence Platform is rules-based and built on published oncology coding standards that are constantly updated. Every decision is explained, every data element is highlighted in the report text for quick validation, and every rule used to select the data element is one click away inside the application. Oncology Data Specialists are no longer just abstracting data elements and doing manual casefinding, and they need powerful, consistent, proven tools that match their workflow.
Automation of Delayed Data Negatively Impacts Outcomes and Value
Technology is a great tool when it works, when the risk is minimal, and when repetitive manual tasks can be safely and accurately automated. But what is the value of automation in oncology if you can't get the proper diagnosis, staging or other critical data for navigation, research, tumor conferences or clinical and operational decisions for weeks or months? That is how LLMs work. To be accurate, they need an International Classification of Diseases (ICD-10) code and staging details confirmed by billing and coding first. Otherwise, the ODS will be changing more data fields than they are validating.
Timing matters everywhere, especially in oncology. Most patient leakage happens in the first 30 to 60 days after biopsy. Cancer data that arrives weeks or months later directly harms patient outcomes and both direct and downstream revenue. Inspirata AI's patented NLP-enabled solutions safely deliver accurate results in real time, at report sign-out.
Proven. Accurate. Safe. Trusted.
Those four words are the standard every oncology data solution should be measured against. Here is how the Inspirata AI Oncology Data Intelligence Platform meets them.
- Proven. Our patented, native, hybrid AI/NLP (not generative AI) Oncology Data Intelligence Platform has been deployed for over two decades and is backed by 25+ years with the National Cancer Institute's (NCI) Surveillance, Epidemiology, and End Results (SEER) Program.
- Accurate. It accesses cancer data in real time from source documents with 99% accuracy and is built on oncology coding rules and standards.
- Safe. It is installed in the customer's environment behind their firewall. PHI stays in their system, data is not shared with anyone, and no change to the customer's or their vendor's cyber security insurance coverage is required.
- Trusted. Hundreds of customers rely on it for cancer data casefinding, data abstraction and NAACCR reporting. It is transparent and uses SEER and local jurisdiction reporting rules. The surfaced cancer data is explained, and decisions are supported by our AI/NLP engineers, physicians and cancer registrars.
The Bottom Line
Generative AI is a powerful tool, and it may eventually get there. But at what financial, patient and industry expense?
In the meantime, health systems and leadership cannot afford to gamble with patient data, patient outcomes, staff satisfaction, job security or revenue. Choose what is proven, accurate, safe and trusted. That is the Inspirata AI Oncology Data Intelligence Platform.
So when evaluating AI vendors for oncology, ask yourself: “Does this vendor pass the P.A.S.T. test (Proven. Accurate. Safe. Trusted.)?”