I’d like to challenge the assumption that clinical coding is an ‘easy picking’ for AI. As I argue, it will be a long time before AI enables all care providers to write, type and speak their clinical notes in a consistent matter in the hospital setting.

The sentiment of digital health leaders at the recent Deloitte breakfast seminar at the Australasian Institute of Digital Health’s Health Innovation Conference (HIC) in August which I attended indicated that clinical coding is one of the areas where AI should start to be applied.

While I don’t disagree that it is a candidate for AI, clinical coding has a number of complexities that need to be addressed by AI-enabled solutions.

Clinical coding is the process of assignment of codes that reflects clinical documentation following the application of the conventions, standards and rules used with a health classification system.

In many countries, including Australia, clinical coding is used to classify and represent admitted patient care and is undertaken in all public and private hospitals. Australia uses ICD-10-AM (International Statistical Classification of Diseases and Related Health Problems, Tenth Revision, Australian Modification), ACHI (Australian Classification of Health Interventions) and ACS (Australian Coding Standards) together as the clinical coding system.

The quality of clinical coded data is important for its multiple uses, including safety and quality monitoring, health research, epidemiology, health care planning, health service evaluation, funding models such as activity-based funding and use in national statistics reporting.

You might be thinking, how hard can it be? You get AI to identify key concepts in clinical documents, apply some rules to assign codes and voila, your clinical coding is done!

However, before we run to replace our in-demand clinical coding workforce with a plug and play AI solution, let me unpack the complicating factors of clinical coding that AI needs to address.

1. Capturing the full scope of clinical codes required for an admitted episode of care

For a single episode of care, the codes assigned need to capture diagnoses, conditions present on admission that impact the care for this admission, medical treatment, surgery, complications of care, cancer morphology, relevant personal and family history, relevant social factors, place and cause of injury, intensive care and ventilation.

In addition, supplementary data on whether conditions were present prior to or after admission, neonatal weight, and episode of care details need to be captured at the time of coding in order to assign an accurate AR-DRG (Australian Refined Diagnosis Related Group).

2. Accessing a complete set of electronic clinical documentation

The clinical documentation required to be reviewed for clinical coding includes progress notes, clinical documentation created during the episode of care, diagnostic investigations, operation and anaesthetic records, and discharge summaries.

For AI-enabled clinical coding all of this clinical documentation, often from multiple sources, needs to be in a digitally readable format and accessible at the time of coding. This is obviously a challenge for paper and scanned clinical documentation.

It also needs to deal with clinical documentation that becomes available after the clinical coding is completed to re-assess the coding accuracy.

3. Interpreting clinical documentation, identifying gaps in clinical documentation and prompting clinicians to document increased specificity

Abstraction involves the review and appraisal of clinical documentation to determine if the assignment of relevant code(s) is warranted, or if there is vagueness or conflicting entries in the health care record. Knowledge of medical science and interpretation of clinical documentation is required.

An AI-solution needs to be able to accurately interpret a broad range of clinical documentation content and styles by clinicians and raise clinical documentation queries for resolution by the documenting clinician prior to coding being completed.

As I said, it will be a long time before in a hospital setting, AI enables all care providers to write, type and speak their clinical notes in a consistent matter.

4. Correctly applying the ICD-10-AM classification conventions

The classification conventions within ICD-10-AM need to be correctly interpreted and applied, for example applying the dual classification principles of aetiology and manifestation using the dagger and asterisk convention.

5. Correctly applying the Australian Coding Standards

The Australian Coding Standards consist of 139 (12th Edition) standards that guide the nationally consistent application of ICD-10-AM and ACHI. For an example of how to apply ACS 0002 Additional Diagnoses, refer to this IHACPA Fact Sheet.

6. Correctly applying national and jurisdictional coding advice and the Clinical Coding Practice Framework

The ICD-10-AM/ACHI/ACS classification system is used in conjunction with National Coding Advice, jurisdictional coding advice and the Clinical Coding Practice Framework.

7. Validating the accuracy of clinical coding

Once the clinical coding is completed and the AR-DRG is calculated, assessing if there is any evidence of additional conditions that could increase the severity of the AR-DRG for that patient, based on the AR-DRG logic and the clinical documentation.

8. Adapting to classification changes

Elements of all of the above change and need to be tested and accurately applied.

Up for a challenge, there are a number of seasoned and entry software companies working on AI-enabled solutions to produce highly accurate and timely clinical coding. Meanwhile, we have a highly skilled clinical coding workforce concerned about what the future of AI and clinical coding will look like.

The Health Information Management Association of Australia’s position is that there is certainly a place for AI integration to be applied to produce quality clinical coding, however any AI solution needs to be carefully considered in light of the above aspects of clinical coding.

HIMAA is chairing a national Clinical Coding and AI Industry Taskforce, which is currently drafting the Australian Clinical Coding AI Adoption Guideline to guide the adoption of AI in consideration of the above factors. This is due to be launched next year following industry consultation.

A significantly revised Clinical Coding Practice Framework will be released next week at the National Health Information Management Conference in Melbourne.

Next year, HIMAA is looking to host an AI and Clinical Coding Showcase for vendors to showcase their clinical coding AI solutions to the health information management and clinical coding community.

If you are interested in expressing interest in participating in the Showcase, please fill out this form.

Sallyanne Wissmann is CEO of the Health Information Management Association of Australia.