Gains:
- Ability to understand the components of the audiological evaluation report and produce a consistent, verifiable report draft with artificial intelligence
- Ability to apply the 'write only from the data I give' discipline, which prevents artificial intelligence from adding fabricated findings (hallucinations) to the report
- Being able to distinguish between the diagnosis, recommendation and referral statements in the report and the responsibility of signing the expert/physician approval.
The outcome of the audiological evaluation is often a report: an official document that reaches the patient, family, ENT physician, school, employer, or institution. The report is a text documenting the tests performed, findings and recommendations, for which responsibility is assumed by signature. In this unit, you will learn how to use artificial intelligence safely in producing report drafts; but we will cover why every finding, recommendation, and referral statement in the report requires expert verification and signature. Main principle: AI can produce a consistent, orderly draft of a report; But it should only write from the real data you provide, should not make up any findings, and the final text should be valid with the approval and signature of the expert/physician.
Components of an audiological report
A typical audiological evaluation report consists of the following sections: It may vary depending on your institution's format:
Identity and context: Patient information (in the institutional system), date of evaluation, referring unit, reason for referral.
History: Related complaints, noise exposure, medication, family history, previous device use.
Test findings: Audiogram summary (type, grade, configuration), tympanometry, acoustic reflex, speech tests, OAE/ABR if necessary. These are based on measurement.
Evaluation and interpretation: The combined meaning of the findings; Type/degree of hearing loss.
Recommendation and plan: Device evaluation, ENT guidance, follow-up, patient education.
Approval: Expert's name, title, signature and date. The signature is where responsibility is assumed.
Caution: The most dangerous mistake in a report is writing the result of a test that was not performed. The AI may add a finding you did not provide (e.g., “acoustic reflexes are normal”) so that the report “looks complete.” This is a hallucination and if signed, it will create serious liability.
To whom does the report go: language by reader
An audiological report is not written to a uniform reader; The language and level of detail varies depending on who you go to. The report to an ENT physician may be technical; The physician knows audiogram terminology and expects diagnostic guidance. In the report to a school or employer, personal clinical detail is minimized, sharing only necessary functional information (e.g., “recommended to sit at the front of the classroom and FM system evaluation”); Medical details are limited for privacy reasons. The summary given to the patient or family is written in plain language. AI can quickly produce these different versions from the same set of verified findings — but the rule of prohibition of fabrication and expert approval applies to each version.
This gives rise to the “one source, many versions” principle in reporting: findings are expertly validated once, then formatted for different readers. While AI facilitates this formatting, it is the expert's responsibility to ensure that the source itself (the measured findings) has not changed and that no new "findings" have been added to any version. Especially in reports sent to third parties (school, employer, insurance), every information shared should be checked in terms of KVKK to see whether it is really the information that needs to be shared.
Prevent hallucination: "write only from the data I give you"
The biggest risk of AI in reporting is fabrication. The model tends to “complete” an incomplete report and may fill in the space you left blank with a plausible but untrue finding. The way to avoid this is to make three rules clear in the prompt:
- Just write from the data I give you. Don't go beyond the raw findings.
- Flag missing data, do not make up. If a test was not performed, write "[no test]" or "[no data]".
- Limit the comment to "possible" language. Establishing a definitive diagnosis sentence; Subject the recommendation to expert approval.
And each report is compared to the source data before signature: does each finding in the draft actually correspond to the data you measured?
Layers of responsibility
Report item
Who verifies?
Why
Test findings
Expert doing the measurement
Calibrated data should not be fabricated
Comment/rating
audiologist
Clinical evaluation
diagnostic statement
Audiologist/physician
diagnostic authority
Referral/routing
Audiologist/physician
medical decision
signature
competent expert
Legal liability
Step by step: report drafting with artificial intelligence
- Structure the raw data. Prepare findings in an anonymous, orderly format (section by section).
- Ban fabrication. Give the rule "Write only from the data I give you, mark the missing".
- Format it. Present the organization's report sections as templates.
- Take the draft and compare it with the source. Check each finding against the raw data.
- Moderate comments and suggestions. Is there a definitive diagnostic language, and is the recommendation subject to expert approval?
- Sign/approve. The specialist/physician approves the final text; An unsigned report is invalid.
Weak prompt / Strong prompt
Weak prompt:
Write a full audiology report for this patient, complete all tests.
“Fill all tests” invites the model to make up tests that have not been done—the most dangerous command.
Powerful prompt:
Your role: report drafting assistant. RULES: write only the findings I give below; DO NOT ADD any tests/results that I have not provided; write missing section[no data]; do not make definitive diagnostic statements, use "possible" language; Connect the recommendation as "according to expert evaluation". Findings (anonymous):- Audiogram: right moderate sensorineural, left within normal limits- Tympanometry: both ears Type A- Acoustic reflex: [not performed]- Speech test: [no data] Format: History / Findings / Evaluation / Recommendation sections. Task: write a draft report from this data. Signature and approval belong to the expert.
A strong request prohibits fabrication, requires the missing to be marked, limits the diagnostic language, and leaves the signature to the expert.
three mini cases
Case 1 — Fabricated capture. An audiologist asks for a draft report, but he has no acoustic reflexes. AI adds "acoustic reflexes within normal limits" to the draft. While comparing it with the source data, the expert realizes that this line is not based on any measurements and deletes it, replacing it with "[acoustic reflex not evaluated]". The hallucination was caught before the signature.
Case 2 — Consistent outline. An audiologist feeds his regular findings (audiogram, tympanometry summary) to the AI with a no-fabrication rule. AI converts these into a clean, house-format draft. The expert verifies each line, makes a single change (softening the recommendation sentence to "based on expert judgment") and signs it. Report writing is reduced from 25 minutes to 8 minutes, accuracy is maintained.
Case 3 — Diagnostic language correction. The artificial intelligence writes "the patient definitely has noise-induced hearing loss" in the draft. The specialist knows that the audiogram alone does not confirm the cause; corrects the statement to say "findings may be consistent with noise exposure; should be interpreted in conjunction with history and clinical evaluation." The language of certainty gravitates towards responsible clinical language.
Copiable prompt templates
REPORT DRAFT (MAKING PROHIBITED) TEMPLATEYour role: report drafting assistant. Write only the findings I gave; do not add any test/results I did not give; write missing [no data]; establish a definitive diagnosis; base the recommendation "based on expert assessment". Findings: [data] Format: [chapters]
SOURCE COMPARISON TEMPLATEMatch EACH finding row in the report draft below with the raw data I provided. Mark any sentence that has no equivalent in the raw data. Draft: [draft]Raw data: [data]
DIAGNOSTIC LANGUAGE SOFTENING TEMPLATEFind the definite/overly assertive statements in the report below ("definitely", "are") and translate them into clinically appropriate "possible/may be compatible/should be evaluated" language. Text: [report]
RECOMMENDATION/REFERRAL CONTROL TEMPLATE In the recommendation and referral section of the report below, mark each statement requiring expert/physician approval and check whether the phrase "with expert/physician evaluation" has been added. Text: [report]
Common mistakes
- It means "fill out all the tests". Pushing the model to fit untested tests.
- Not comparing with the source. Signing the draft with raw data without checking it line by line.
- Leaving the definitive diagnosis language. Avoid using expressions such as "there is/definitely" into clinical language.
- Not subjecting the recommendation to expert approval. Presenting dispatch/plan statements as automatic.
- Using without signature/approval. Distribute the draft as an official report.
In summary
The audiological report is an official text in which the findings based on measurement and expert interpretation are documented with a signature. AI can produce a draft of this report in an orderly and consistent manner; But they should only write from the real data you provide, should not make up any findings, and should not use definitive diagnostic language. Each draft is compared with the source data, diagnosis and recommendation language is checked, and the final text is validated with the signature of the specialist/physician. An unsigned report is invalid.
Application task
Prepare a findings set with anonymous, incomplete information (some tests "[not performed]"). Get a draft with the "report draft (fabrication prohibited)" template, then deliberately relax the rule and say "fill out all tests" and get a second draft. Compare the two outputs: is there a spurious finding in the second? Flag any lines that do not match the "source comparison" template and apply the "diagnostic language smoothing" template.
checklist
- [ ] I have provided the raw findings anonymously and structured.
- [ ] I made a no fabrication rule and a "[no data]" rule.
- [ ] I compared the draft line by line with the raw data.
- [ ] I have changed the language of definitive diagnosis to the language of clinical "possible".
- [ ] I have subjected the recommendation/referral statements to expert approval.
- [ ] I approved the final report with the signature of the specialist/physician.