For engineering teams building ambient AI workflows, it is remarkable how quickly these capabilities have evolved from a “nice-to-have” feature to an expected part of modern clinical practice. As vendors have rolled out listening tools at scale, the baseline experience – capturing a visit, producing a transcript, and drafting a clinical note – has become widely available and increasingly similar across products. This shift is good for adoption, but it also means the market is entering a phase where organizations can no longer rely on transcription or summarization alone as their primary differentiator.
The specificity gap in ambient AI
The hard part of ambient AI is not producing fluent text. It’s producing clinical meaning with enough precision to be reused. A conversation can be accurately captured and even summarized, yet still fail to support coding, quality reporting, prior authorization, or decision support because the output is not normalized, not specific, and not aligned to clinical terminologies.
For example, “knee pain” may be a reasonable narrative summary, but it does not distinguish between laterality, acuity, cause, associated findings, or whether it represents a symptom versus a diagnosis. And when ambient outputs remain generic, organizations end up redoing work downstream – manual review, coder queries, and back-and-forth clarifications – negating much of the time savings that ambient documentation promised in the first place. Ultimately, clinical specificity is not being found, and real money is being left on the table.
Accelerating ambient listening with structured data
IMO Health has developed an Ambient Listening solution accelerator that converts medical transcripts into structured clinical data using IMO IDs – unique identifiers backed by 30 years of curated clinical knowledge – enabling a high degree of specificity for each clinical term.
The following steps outline the end-to-end workflow implemented in the Ambient Listening solution accelerator.
Step 1: From transcript to SOAP note
The accelerator starts with uploading a standard patient-clinician transcript to the application. This transcript can be from any therapeutic area, such as cardiology, nephrology, mental health, etc. Once the transcript is uploaded, we use the Amazon Bedrock Nova Pro model to convert the text into a SOAP note. This categorizes the content of the transcript into the sections of Subjective, Objective, Assessment, and Plan, thus preparing a full clinical note.
Step 2: Extracting and normalizing clinical entities
Next, we run natural language processing (NLP) on the SOAP note using the IMO Entity Extraction API. Using named entity recognition (NER), the IMO Entity Extraction API successfully identifies all the problems, procedures, medications, and labs in the note.
Then, in the same step, the magic happens. Using the IMO Precision Normalize with Enrichment API, we assign each clinical entity its own precise IMO ID. For example, type II diabetes mellitus with retinopathy and macular edema, right eye would receive the IMO ID 43209870.
The IMO ID is foundational to the specificity workflow. Each normalized clinical concept is mapped to a proprietary semantic data layer developed and curated by IMO Health clinicians over decades. This enables clinical entities to be represented with a high degree of precision and mapped to over 30 different medical code systems, such as ICD-10-CM, SNOMED CT®, LOINC®, CPT®, and HCPCS. This level of specificity helps support downstream workflows such as coding, reimbursement, and quality reporting.
Step 3: Refining clinical terms with the Knowledge Graph
The last step in the solution accelerator is the diagnostic specificity workflow. Here, we use the IMO Core Search API to leverage IMO Health’s Knowledge Graph for each clinical entity. This allows us to utilize the IMO ID to arrive at the highest degree of clinical detail. For each term that can be additionally specified, the user has the agency to select the appropriate modifiers. For example, if the term “chest pain” is present, the application would highlight if it could be further refined. All possible refinements of this term, such as the type of chest pain and its location, would then be available for selection.
This is entirely done using the power of the Knowledge Graph, as each base term can be refined using its child codes. The base term is represented in the form of a “node” in the graph and is connected to related nodes using an “edge.” In this case, “chest pain” is the base term, and a possible child code could be “angina pectoris with coronary microvascular dysfunction.” Both of these terms are represented with nodes and are interconnected with edges.
Using the precision of the IMO ID, the child term is then mapped to its ICD-10-CM and SNOMED CT codes. This process is repeated for every clinical entity identified in the note, with each code grounded in clinical evidence. After each code is maximally specified, the enriched codes can be downloaded in JSON format and, in an actual clinician workflow, sent for billing and reimbursement.
Turning precision into reimbursement
The Ambient Listening solution accelerator helps generate highly specific clinical codes within the clinician workflow, enabling more accurate documentation and downstream reimbursement processes. No matter which code system a clinician uses – ICD-10-CM, LOINC, CPT, etc. – clinical specificity is always the key. This has been the driving force behind IMO Health for 30 years, and it’s what motivates us to continue to meet the needs of clinicians as they adopt ambient AI as a central system of record within their workflows.
