This post is part of Pattern Labs, our ongoing research series focused on translating real-world experimentation into defensible, production-ready workflows for mass tort litigation. The previous two installments covered why large language models have a structural tendency to return positive findings and why a single prompt cannot reliably evaluate a case. This installment examines what happens when a model encounters the kind of clinical language that appears constantly in mass tort records: ambiguous, layered, and rarely written with litigation criteria in mind.
Clinicians write medical records for other clinicians, in the shorthand of a specialty, an institution, and sometimes a single provider's documentation habits. A general-purpose AI evaluating those records for litigation purposes has none of that context. It reads what's on the page and returns what it finds.
The straightforward cases take care of themselves. Judgment is required when something appears in the record, but its meaning isn't obvious without clinical and litigation-specific expertise.
Paraquat litigation, which centers on Parkinson's Disease claims, offers a clear example of how far that gap can stretch. Consider how "parkinsonism" appears in those records. The term shows up constantly, and it means different things depending on where it appears and how it is used.
Early in a diagnostic workup, a provider might document parkinsonism to describe observed motor symptoms before a definitive diagnosis has been made. Later records for the same patient may contain a confirmed Parkinson's Disease diagnosis. In other files, parkinsonism appears as a catch-all descriptor for a set of symptoms that never resolve into a qualifying diagnosis. In still others, it refers to a condition caused by medication side effects or head injury, which falls into non-qualifying territory entirely.
A general-purpose model reading the word "parkinsonism" in a record has no mechanism to distinguish between those scenarios. It sees a term associated with the relevant condition and registers it as a finding. Whether that finding reflects a confirmed diagnosis, a documented symptom pattern consistent with a developing condition, or a non-qualifying cause requires clinical knowledge and litigation-specific guidance that the model doesn't have unless it has been explicitly built in.
Getting this right in Paraquat required working through the distinction with medical reviewers and litigation experts who understood both the settlement criteria and how neurologists actually document these conditions over time. The result was evaluation logic that treats a confirmed PD diagnosis, qualifying symptom patterns, and non-qualifying parkinsonism variants as categorically different findings rather than variations of the same signal. That logic doesn't transfer automatically to the next litigation, and it isn't something a general-purpose model arrives at on its own.
The parkinsonism example illustrates a pattern that surfaces in every litigation Pattern has worked across. Medical language is imprecise by design. Providers document what they observe, what they suspect, and what they want to monitor, often in the same record and sometimes in the same sentence. Settlement criteria require the reviewer to draw hard lines between those categories.
A nursing note that reads "monitor for signs and symptoms of DVT" looks like a condition mention. Clinically, it reflects routine prevention protocol, not provider suspicion of an existing condition. An unconfirmed diagnosis requires active clinical suspicion, language like "possible," "suspected," "rule out," or "differential," tied to symptoms or an active evaluation. A model that can't draw that line will generate findings from boilerplate prevention language, producing false positives at scale across every case where that documentation pattern appears.
This is standard clinical documentation. The specialties most commonly represented in mass tort records, neurology, pulmonology, oncology, write this way routinely. Building evaluation logic that handles it correctly requires working through real records with people who know both the clinical conventions and the settlement criteria well enough to define exactly where the line falls.
Most AI tools are evaluated on whether they got the answer right. The question litigation managers are actually asking is how fast their team can move through a file once the answer arrives. That's why explainability is built into how Pattern structures outputs.
A finding that arrives with a direct reference to the specific record language supporting it takes seconds to verify. A finding that arrives as a conclusion without that context requires the reviewer to go back into the record and confirm it from scratch. At the volumes mass tort requires, that difference in verification time is where review capacity goes.
Ryan Schmitt is a Product Manager on the AI team at Pattern Data.
Why does general AI misread clinical language in medical records?
Clinicians write records for other clinicians, in the shorthand of a specialty, an institution, and sometimes one provider's habits. A general-purpose model has none of that context. It reads what's on the page and returns what it finds, so a term associated with a qualifying condition registers as a finding whether or not the record actually supports one.
How can the same medical term mean different things in one litigation?
In Paraquat records, "parkinsonism" can describe observed motor symptoms before a diagnosis, a confirmed Parkinson's Disease diagnosis, a symptom pattern that never qualifies, or a non-qualifying cause like medication side effects or head injury. The meaning depends on where and how it appears. Pattern's evaluation logic treats those as categorically different findings, not variations of the same signal.
Why does explainability matter as much as accuracy in case evaluation?
Getting the answer right is only half the job. What litigation managers actually need is speed of verification. A finding that arrives with a direct reference to the record language supporting it takes seconds to confirm. A finding delivered as a bare conclusion sends the reviewer back into the record from scratch. At mass tort volume, that difference is where review capacity goes.