Enhancing Inpatient CDI Outcomes: Best Practices in Documentation and Coding

Last October, I led an AHIMA25 workshop titled “Enhancing Inpatient Clinical Documentation Integrity (CDI) Outcomes: Best Practices in Documentation and Coding.” The session used case-based scenarios to examine five high-risk inpatient conditions: sepsis, acute kidney injury (AKI), end-stage renal disease (ESRD), congestive heart failure (CHF), and malnutrition.

The objective was not only to revisit definitions, but also to evaluate how documentation and coding decisions withstand clinical validation, audit, and denial scrutiny. We also examined how artificial intelligence (AI) changes both the opportunity landscape and the risk profile of CDI work.

Recognizing Automation Bias

Two moments, in particular, reframed our discussion. First, the workshop answer key was intentionally designed to mirror real-world vulnerabilities. It paired sound industry concepts with a small number of inappropriate “final-coded” diagnoses generated by AI and missed by the facility health information (HI) teams. The aim was neither to sensationalize nor vilify the unfortunate outcome of the technology, but rather to demonstrate how easily an authoritative-sounding output can potentially dull skepticism to compromise the “pause and verify” step that experienced coding, CDI, and clinical teams depend on.

In our scenario, the erroneous AI-generated code proposals and query suggestions were misinterpreted by a novice CDI team member and a fairly experienced coding manager. Part of the reason for this was a pervasive provider progress note error due to misuse of another AI tool, Computer-Assisted Physician Documentation (CAPD), by the discharging provider, with an overt misclassification of the heart failure type. Some workshop attendees seemed genuinely surprised, others less so.

This is the emerging warning for the HI workforce and clinicians: when AI-generated language appears complete and confident, it can quietly substitute pattern recognition for clinical reasoning. In the coming years, the competitive advantage will not be access to AI itself. It will be the ability to intelligently govern the human–AI interface strategically. In doing so, these splendid and powerful AI insights might also support compliant, defensible decisions rather than accelerating ambiguous, unsupported, or prematurely definitive conclusions.

Governing Five High-Risk Conditions

The technical core of the workshop focused on five conditions that frequently trigger clinical validation denials. For each condition, we reviewed a commonly encountered documentation flaw and the corresponding risk introduced by automation.

Sepsis: The Discipline of Causality

Sepsis documentation often fails because it does not clearly express causality, not because the underlying clinical science is misunderstood. A recurrent denial driver is missing linkage, in which a list of findings substitutes for a defensible clinical chain of reasoning. Sepsis depends on explicit relationships among infection, organ dysfunction, and the management decisions that follow. When the record does not clearly establish those relationships, the narrative becomes vulnerable under external review, even when the clinician’s intent appears obvious to the bedside team.

The AI-related risk for sepsis is not limited to an overtly “wrong diagnosis.” A more frequent failure is polished documentation that appears clinically complete while still lacking linkage or clear provider attribution. This risk is amplified by copy-forward behaviors, in which tentative language can be repeated until it is treated as established fact. Query standards and technology standards converge on the same operational expectation: AI can help teams identify when critical elements of the clinical story are missing, but it must not supply the story on the provider’s behalf.

Acute Kidney Injury: Establishing the Baseline

AKI is frequently validated or denied on one central question: what is the patient’s baseline, and what changed because of the renal injury? In practice, the record should communicate baseline awareness, trajectory, suspected contributors, and concrete management responses (for example, medication adjustments, hemodynamic strategy, fluid management, avoidance of nephrotoxins, or consultation decisions). Vulnerability arises when documentation reflects laboratory trend awareness without demonstrating the clinical reasoning and the actions triggered by the kidney injury.

For AKI, the AI-related risk is often premature certainty. Automation can convert a differential diagnosis into a definitive causal statement that outpaces the clinician’s documented workup, and that “hardening” can then spread through the record via templated carry-forward documentation. Technology standards emphasize that automated identification of opportunities must remain grounded in clinically supportable indicators and must not drive leading clarification.

ESRD: Documenting Dependency, Not Labels

ESRD is not a laboratory impression; it is a dependent clinical state. The record must make dialysis status unmistakable, including modality, schedule, access, and what was managed during the admission. A frequent clinical validation failure is “unsubstantiated diagnostic persistence,” in which advanced chronic kidney disease is labeled “end-stage” as a shortcut without confirming ongoing dialysis dependence or transplant status. That shortcut can convert a complex but supportable renal story into a denial target.

One AI-related risk for ESRD is inertia. Once a label enters the record, automation can reinforce it because the system may treat prior language as a reliable signal. For high stakes diagnoses like ESRD with implications to disability status, documentation should reflect verification steps that connect the diagnosis to orders, treatment, and accountable provider assessment rather than inherited phrasing.

CHF: Specificity as a Clinical Narrative

CHF documentation most often fails when it omits acuity, type, or the decompensation narrative. Heart failure is a clinical syndrome that should be described with sufficient specificity to support what the clinician believed, why that conclusion was reached, and what actions followed. Evidence is often distributed across imaging, oxygen requirements, diuretic intensity, weights, volume status examination, and response to therapy. A defensible record consolidates those signals into a clear provider-owned conclusion that aligns with the plan of care and minimizes ambiguity that can be challenged in retrospective review.

Given the above-mentioned provider misuse of CAPD that erroneously classified CHF type in one of our case studies, AI-related risk includes definitive language that the clinician expressed in the electronic health record (EHR), but on retrospective reflection never truly intended. Overconfident automation can broadcast a fragile clinical narrative that reads coherently but lacks the documented reasoning and attribution required to withstand clinical validation.

Malnutrition: The Requirement for Provider Ownership

Malnutrition remains a common example of “evidence without durability.” Organizations may document abundant objective indicators and still lose the diagnosis because the provider does not demonstrate ownership within the medical decision-making narrative. The provider, rather than the registered dietitian (RD), must document that malnutrition is clinically meaningful to the encounter. Under clinical validation, a durable record reflects provider acknowledgment and integration into the plan of care, even when registered dietitian assessments are robust.

The AI-related risk is provider duplication of the dietitian note that reads as templated repetition rather than clinician judgment. Certain providers ostensibly lack confident familiarity with evaluations and assessments that meaningfully incorporate ASPEN malnutrition criteria, and instead mirror the exact phrasing in RD assessments across their own notes without showing how the diagnosis informed decisions. This undermines clinical accuracy and credibility.

Operationalizing Strategy: Integrity and Standards

Across these diagnoses, outcomes improve when documentation is treated as evidence and reasoning rather than a memory aid. That orientation aligns with compliant query expectations: clear clinical indicators, neutral language, and queries designed to clarify provider intent rather than steer it. Documentation must also satisfy broader medical record requirements that support payment and withstand external review.

The query process is the point at which best practices become operational. High-performing programs convert ambiguity into standards, including shared thresholds for when a query is warranted, consistent conventions for presenting clinical indicators, reliable escalation pathways for clinical validation concerns, and ongoing auditing of query quality. In this context, AI can assist by surfacing patterns, prioritizing review, and supporting workflow mechanics. However, industry standards are explicit that governance remains human work: clinicians and CDI professionals must control the final wording, timing, intent, and compliance posture of clarification.

AI increasingly scales up technological insights that are crucial and welcomed in CDI and coding production, and it does not resolve documentation ambiguity by default. Disciplined oversight is a mandate. Without such oversight, AI might also scale up ambiguity faster than teams can contain. Technology standards and query practice standards point in the same direction: the path forward is governance that regulates the optimism with a note of caution that still respects the power of AI. Organizations that succeed will be those that design AI-enabled workflows to elevate clinical reasoning, maintain provider accountability for the clinical story, and treat documentation integrity as an operational discipline rather than an afterthought.

Tarman Aziz, MD, CCDS, is CEO and Founder of CDIQ Consulting in Clearwater Beach, FL.