Units
1. Introduction to Artificial Intelligence in Audiology: Roles, Boundaries, Validation and Ethics 2. Audiogram Reading and Comment Support: Validating AI Draft 3. Pre-Evaluation of Hearing Screening Results: Newborn, School and Workplace 4. Tympanometry, Acoustic Reflex and Middle Ear Tests Interpretation Support 5. Hearing Aid Selection and Suitability Assessment 6. Hearing Aid Programming (Fitting) Workflow and Real Ear Measurement 7. Artificial Intelligence in Tinnitus and Balance/Vestibular Assessment 8. Patient and Family Education: Plain Language and Material Production 9. Clinical Reporting and Documentation: From Draft to Expert Approval 10. Device Data, Teleaudiology and Privacy (KVKK) 11. End-to-End Case: AI-Powered Workflow from Scan to Follow-Up
Unit 2 / 11

Audiogram Reading and Comment Support: Validating AI Draft

Gains:

  • Ability to interpret the basic components of the audiogram (dB HL, frequency, air conduction/bone conduction, masking) and the type and degree of hearing loss by verifying it with a calibrated measurement, not with an artificial intelligence sketch
  • Ability to provide artificial intelligence with audiogram data in a structured format, produce a safe preliminary interpretation draft, and translate the diagnostic language into expert language
  • Ability to catch common audiogram interpretation pitfalls (masking dilemma, air-bone gap, configuration blindness) before the prompt

The audiogram is the document that an audiologist looks at most. The audiogram is a graph that plots the patient's hearing thresholds (the lowest sound intensity they can hear) by frequency: on the horizontal axis is frequency (pitch of sound, in Hertz — Hz, typically 250–8000 Hz), and on the vertical axis is intensity (loudness of sound, hearing level in decibels — dB HL). The graph is read from top to bottom: the lower the threshold (the greater the dB HL number), the greater the hearing loss. In this unit we will learn how to use AI as a rough draft generator for audiogram interpretation, but why you should always verify the final interpretation with calibrated measurement and the clinical eye.

The main principle is this: AI can draft you a neat, readable pre-commentary; However, the type and degree of hearing loss and the need for masking become clear with calibrated device data and expert evaluation. Instead of telling the artificial intelligence to "interpret" the audiogram, the correct approach is to say "a regular outline will emerge from the numbers I give, I will verify it."

Basic components of the audiogram

Let's clarify a few concepts to read an audiogram correctly; You must understand each of them in your own mind before giving them to artificial intelligence.

Airway (HY): Sound enters the outer ear through headphones or speakers and tests the entire auditory pathway (external-middle-inner ear). On the audiogram, it is usually marked "O" for the right ear and "X" for the left ear.

Bone conduction (BO): A vibrating head (bone vibrator) is placed on the bone behind the ear; The sound bypasses the outer and middle ears and directly tests the cochlea (the hearing organ in the inner ear). It is usually marked with "<" and ">" on the audiogram.

Air-bone gap (AAC): The airway threshold is significantly worse than the bone gap. This range suggests that there is an obstacle in the transmission of sound from the outer/middle ear, that is, there is a conductive component (originating from the outer/middle ear).

Masking: When the sound passes through the skull to the other (non-tested) ear while testing one ear, it is called cross-hearing. If the difference between the thresholds of the two ears is large enough to cause this, it is "masked" by introducing noise into the untested ear so that the threshold actually belongs to the tested ear.

Caution: An audiogram where masking is not performed where required may show a threshold in the wrong ear. Artificial intelligence interprets the numbers you give as they are; You must ensure and verify whether masking is performed.

Type and degree of hearing loss

There are two main things the audiogram tells you: the type and degree of loss.

Type tells where the loss is. If the air and bone conduction are poor together and there is no gap, the sensorineural (inner ear/nerve origin) type is considered. If the airway is poor (if there is a gap) while the bone conduction is close to normal, conduction type is considered. If both are together, it is a mixed type.

Grade describes how much loss there is and is usually classified according to average thresholds (normal, mild, moderate, moderate-severe, severe, very severe). The table below provides a common framework; however, rely on the threshold ranges used by your institution and current guidance.

Degree

Approximate threshold range (dB HL)

Possible impact on daily life

normal

≤ 20

No obvious difficulties

lightweight

21–40

Difficulty speaking quietly/distantly

medium

41–55

Difficulty speaking normally

intermediate-advanced

56–70

Need to speak loudly

Next

71–90

Communication without a device is very difficult

too far

> 90

Device/implant evaluation

Artificial intelligence can remind you of this classification; But which frequencies are averaged, the validity of the calibration, and the reliability of the test (patient cooperation, age-appropriate method in the child) are your responsibility.

Step by step: safe audiogram pre-interpretation with artificial intelligence

  1. Structure the data. Table the right/left, air/bone conduction thresholds anonymously, frequency by frequency, in dB HL.
  2. Specify masking status. Add which thresholds are masked; If it is missing, write down "masking should be evaluated".
  3. Limit the role. Tell the artificial intelligence "don't make a diagnosis; just outline the type and degree from the number I give you, don't use precise language."
  4. Get the draft and verify it. Compare the resulting type/degree interpretation with your own reading and device data.
  5. Integrate with clinical context. Integrate with history, otoscopy, tympanometry and speech tests.
  6. Take it to the expert language. Convert the draft into a validated statement appropriate to the report/patient language.

Weak prompt / Strong prompt

Weak prompt:

This patient's audiogram is poor, right ear is around 60. What type of loss device do I need?

Incomplete and unstructured data, direct diagnosis and device decision request, no masking and bone conduction information. AI is forced to make up the gaps.

Powerful prompt:

Your role: assistant drafting pre-interpretation to the audiologist. Making a diagnosis; just DRAFT the type and degree from the numbers I give, use the "possible" language, mark the missing data as [no data]. 1000=30, 2000=45, 4000=50Masking: applied. Left ear: [no data].Task: sketch the possible type (range?) and degree for the right ear, what additional testing is needed for confirmation. State that the decision is up to the audiologist.

Strong prompt gives structured data, sets roles and boundaries, prevents fabrication, and requires verification.

three mini cases

Case 1 — Catching the gap. An audiologist gives the data of HY 1000 Hz = 50 dB, KY 1000 Hz = 15 dB in the right ear to the artificial intelligence. YZ outlines "air-bone gap approximately 35 dB, possible conductive-type component; should be confirmed by otoscopy and tympanometry." The specialist combines this with otoscopy and tympanometry and directs the patient to a middle ear evaluation. AI accelerated the draft, the decision remained with the expert.

Case 2 — Masking warning. An audiometrist enters data for the right ear 4000 Hz = 20 dB, the left ear the same frequency = 80 dB and states that no masking is applied. AI reminds that the difference can lead to crossed hearing, with left ear threshold being unreliable without masking. The specialist repeats the test with masking. The 60 dB difference made visible a trap that could have been overlooked.

Case 3 — Hallucination prevention. An expert enters only the right ear data but says "interpret both ears." The AI ​​tends to fit values ​​for the left ear, which is not given. When the expert adds the "make up missing data, write [no data]" rule to the prompt, the AI ​​leaves the left ear empty. Lesson: if you don't set limits, the model will fill the gap.

Copiable prompt templates

AUDIOGRAM DATA CONFIGURATION TEMPLATE Plot the following thresholds into an anonymous table (frequency row; right HY, right KY, left KY, left KY column). Write missing cells [no data], fitting. Raw data: [thresholds]

TYPE-RADE DRAFT TEMPLATEYour role: pre-comment draft assistant. From the thresholds I have given, for each ear (1) is there an air-bone gap, (2) possible type, (3) possible degree sketch is drawn with the language "possible". Making a definitive diagnosis; Specify additional testing required for verification. Data: [table]

MASKING CONTROL TEMPLATEIs there a difference between the two ear thresholds I have given that requires masking? At which frequencies should masking be evaluated? Give the result as a list; the decision is up to the expert. Data: [thresholds]

TEMPLATE FOR TRANSLATION INTO EXPERT LANGUAGETranslate the draft I have verified below into a single paragraph in measured and "possible" language, appropriate to the clinical report. Adding fabricated findings. Draft: [verified draft]

Common mistakes

  • Bypassing the bone pathway. Just give the airline and ask for a type comment; It cannot be seen whether there is a gap or not.
  • Ignoring the masking dilemma. Accepting the threshold as accurate without masking for large interaural differences.
  • Configuration blindness. Looking only at the average and missing a single frequency notch (e.g. 4000 Hz noise notch); If the AI ​​returns an average, the detail is hidden.
  • Requesting definitive diagnostic language. Telling AI to "diagnose"; Using the draft as a final report.
  • Making up missing data. Allowing the model to fill empty cells when not setting a limit.

In summary

Audiogram; It is based on the concepts of frequency, dB HL, air/bone conduction and masking. The AI ​​can generate a neat type-degree outline from the structured numbers you give it and remind you of pitfalls like masking; but the type, grade and reliability of the test become clear through calibrated measurement and expert evaluation. Provide data anonymously and structured, print the missing [no data], always verify the draft with your own reading.

Application task

Plot the thresholds of an (anonymised) audiogram you have into a table with the "data structuring" template, and get a preliminary interpretation with the "type-degree draft" template. Then do your own independent reading and compare it with the AI's draft: where did you come out the same, where did you come out different? Note whether it misses the masking requirement and convert the final verified comment into a single paragraph with the “expert translation” template.

checklist

  • [ ] I gave the thresholds in an anonymous and structured form.
  • [ ] I established air and bone passage together; I checked the range.
  • [ ] I considered the need for masking.
  • [ ] I marked the missing data as [no data], preventing fabrication.
  • [ ] I validated the AI ​​manuscript with my own reading and clinical context.
  • [ ] I translated the last comment into expert language, with measured, "possible" language.
Previous Next