Coding Automation

Under the Patient-Driven Groupings Model (PDGM), the primary diagnosis assigns the episode to one of 12 clinical groupings and secondary diagnoses drive the comorbidity adjustment (none, low, or high), so coding directly determines payment. Coding also gates the revenue cycle: the OASIS cannot be finalized and the Notice of Admission and claims workflow cannot proceed cleanly until diagnoses are settled. Many agencies outsource coding at a per-chart fee with turnaround measured in days, which delays everything downstream. Automation attacks both the cost and the delay by producing a coded draft within minutes of documentation being available.

Current tools use language models to read unstructured sources, including faxed referral packets, hospital discharge summaries, medication lists, and the clinician's assessment, then propose a primary diagnosis and sequenced secondary diagnoses with supporting evidence linked to the source text. Good systems also flag problems rather than papering over them: symptom codes that are unacceptable as primary diagnoses under PDGM, diagnoses lacking supporting documentation, and conflicts between the referral and the assessment. Evidence linking matters most, because a reviewer who can see why a code was suggested can validate it in seconds instead of re-reading the chart.

Official coding guidelines require code assignment to be supported by physician documentation, and responsibility for claim accuracy stays with the agency regardless of what software produced the draft. A practical control structure:

• A credentialed coder or trained clinician reviews and approves every AI-coded chart, at least until measured accuracy justifies risk-based sampling
• QA audits a sample of automated coding monthly, tracking agreement rates and payment impact of corrections
• Physician queries are generated when documentation does not support a clinically suspected diagnosis
• Upcoding risk is watched specifically, since errors that inflate comorbidity adjustments create overpayment exposure

Ask vendors for accuracy measured against credentialed-coder consensus on home health charts, not generic clinical documents, and ask how accuracy is tracked in production. Probe the failure modes: what happens with a 40-page referral packet of poor-quality faxes, or a patient with 15 active diagnoses? Confirm the tool sequences codes under home health rules, including PDGM-acceptable primary diagnoses, rather than just extracting every code mentioned. Finally, look at workflow fit: suggestions that land inside the EHR coding screen with evidence citations get reviewed properly, while output delivered as a separate document gets rubber-stamped.