![\"Mother](\"https:\/\/rtmedical.com.br\/wp-content\/uploads\/2026\/08\/laudo-radiologico-familia-pediatria-scaled.jpg\")
How the study was done<\/h2>\n
The authors made an interactive, lay-oriented version of every<\/strong> pediatric imaging report available through the patient portal at both institutions, for exams performed between January 2024 and November 2025. Conversion was handled by a commercial natural language classification algorithm (Scanslated Inc., Tampa, Florida) that turns the standard report into a patient-centered web page, with technical terms annotated and findings made navigable.<\/p>\n Families were then invited to complete a user-experience survey. It is worth being precise about the design: this is not a randomized trial with a clinical endpoint, but a real-world assessment of use and perception based on voluntary responses. What the study measures is self-reported comprehension and satisfaction — legitimate, measurable outcomes, but different from “the family made better clinical decisions.”<\/p>\n The researchers’ rationale is blunt: “Terminology in radiology reports can be difficult for patients and families to comprehend, which may cause anxiety and impact the ability of patients and families to make informed decisions about the care they receive.” In pediatrics that statement carries double weight: the reader is not the patient but a caregiver, and the decision that follows almost always involves more than one adult.<\/p>\n The radiology report was never written for the patient. It is a technical document addressed to another physician, with differential diagnoses, incidental findings, deliberate hedging (“cannot be excluded”) and Latin vocabulary. For decades that produced no friction, because the text reached the patient filtered through the ordering physician.<\/p>\n That filter is gone. Information-blocking rules now require immediate release of results into portals, and today a large share of reports is opened by the patient before any clinical conversation. We have covered the cruelest edge of this shift, when patients learn of a cancer diagnosis through the portal<\/a>, alone, on a weekend. A plain-language report does not solve the problem of bad news without a mediator, but it removes the layer of noise stacked on top of it.<\/p>\n A technical aside is warranted, because coverage tends to blur the two. The tool used in the study is a classifier: it recognizes known entities and expressions in the report text and maps them to pre-written, validated explanations, assembling an interactive page. The behavior is deterministic and auditable — the same sentence always yields the same explanation, and no new content is invented.<\/p>\n A generative language model works differently: it rewrites the report by producing new text, with far better fluency and coverage of cases no dictionary anticipated — at the cost of hallucination risk and run-to-run variation. Both approaches have supporting evidence: we have reported that LLMs generate simpler versions of radiology reports<\/a> and that GPT-5 improves patient comprehension of reports<\/a>. Choosing between classifier and generative model is, in practice, a risk management decision: predictability versus flexibility.<\/p>\n Operationally, the most useful lesson is that the translation layer was applied to all<\/em> exams, not offered as an opt-in feature on request. That matters because the family that most needs the explanation is precisely the one least likely to know it exists. Enabling it by default is an equity decision, not just a usability one.<\/p>\n It is also relevant that none of this requires changing the signed report. The radiologist keeps producing the same technical document, with the same legal responsibility; the lay version is a derived artifact generated downstream, in the portal. That separation is what makes the model deployable without renegotiating workflow or adding to turnaround time — a sensitive point in departments already running at capacity, as the decade-long rise in report turnaround times<\/a> makes clear.<\/p>\n Outside English-speaking markets, two obstacles show up before the technology does. First, many services still deliver results as a generic link or PDF, with no authenticated patient portal, which blocks the interactive layer entirely. Second, the problem is linguistic: classification tools depend on term dictionaries, and the mature dictionaries are in English. Porting that vocabulary is clinical curation work, not machine translation — “ground-glass opacity” and its literal translation are only equivalent to someone who already knows what both mean.<\/p>\n The constraints are the expected ones for a user-experience study. The sample is self-selected: those who answered the survey are likely those who opened and used the feature, which tends to inflate positive perception. Comprehension is self-reported rather than tested with objective questions about report content. And the data come from two US academic centers whose socioeconomic profile and digital literacy do not represent the average.<\/p>\n The next rung of evidence is measuring outcomes rather than perception: does the accessible version cut calls to the department, reduce no-shows for follow-up imaging, shorten time to the return visit? Those metrics already live in the RIS and the scheduling system, and they would turn “families liked it” into a management argument. In the meantime, the Stanford and Duke finding already supports a defensible decision: if the patient is going to read the report anyway, it is worth offering a version built for them.<\/p>\nWhy report language became a clinical problem<\/h2>\n
Natural language classification is not the same as an LLM<\/h2>\n
What it changes for the radiology department<\/h2>\n
Limitations and next steps<\/h2>\n