Why It’s Time to Rethink ICD Coding: A Physician’s Perspective on Fairness, Burnout and Better Technology

By: Steven L. Sivak, MD FACP
June 23rd, 2025

In the fast-paced world of inpatient medicine, physicians are constantly juggling clinical care, documentation, communication, and patient advocacy. Amid these responsibilities, one particularly frustrating task has grown in importance but not in clarity: the assignment of precise ICD-10 diagnosis codes. These codes not only define a patient’s condition—they also directly affect hospital reimbursement through the Medicare Diagnosis-Related Group (DRG) system.

Unfortunately, the language of ICD-10 often doesn’t align with the way physicians think or document. This disconnect not only complicates our workflow, it contributes to cognitive overload and burnout, while also penalizing hospitals for inaccurate coding—even when excellent care is provided. Let me explain the problem, and then offer a path forward.

Take the term "urosepsis"—a perfectly acceptable and frequently used diagnosis by clinicians to describe a systemic infection stemming from a urinary source. Yet, in the world of ICD-10, “urosepsis” is considered non-specific and often rejected or downcoded by hospital coders or auditors. Instead, coders look for language like:  Sepsis due to Escherichia coli (E. coli) (A41.51)  Severe sepsis without septic shock (R65.20)

Each of these codes maps to a higher-weighted DRG, which in turn leads to higher reimbursement for the hospital. But for that to happen, the physician must explicitly document these terms—sometimes even when they’re not the natural way a seasoned clinician would phrase it.

Hospitals invest heavily in physician education, “query” processes, and Clinical Documentation Improvement (CDI) teams to nudge physicians toward using the exact right phrases. These teams comb through charts looking for supporting data—vital signs, lab results, cultures—that could justify a more specific diagnosis code. Then they send queries back to the physician: “Can you clarify whether the patient had sepsis due to a specific organism?”

While the intention is good the execution, however, places an increasing documentation burden on physicians, many of whom are already reporting record levels of burnout. According to a 2022 survey published in JAMA, nearly 63% of physicians reported at least one symptom of burnout, with documentation and administrative burden being top contributors.

Hospitals don’t want to leave money on the table. CMS (Centers for Medicare & Medicaid Services) payment systems are increasingly data-driven, and ICD codes are

the engine behind both reimbursement and quality metrics. But tying hospital payments and quality scores to specific words—when those words may not align with standard clinical language—creates a misaligned system where financial viability depends on documentation phrasing, not quality of care.

The burden of this documentation strategy ultimately falls on the bedside physician, who is already responsible for diagnosing, treating, educating, and documenting all while managing clinically complex patients.

And while CDI specialists are helpful, the reality is this: the EMR already contains the necessary data to assign more specific codes. It's just not being leveraged intelligently. We don’t need to overhaul the entire ICD-10 system to fix this. What we need is an intelligent intermediary—a system that understands both clinical language and coding language, and can translate between the two without asking more from the physician. Imagine this workflow:

1. The physician writes: “Urosepsis suspected. Positive urine culture for E. coli. Blood cultures pending. Patient febrile, tachycardic. Started on ceftriaxone.”

2. An AI-enabled coding tool, trained on both ICD coding logic and medical documentation styles, reads this note.

3. It maps the clinical data to the correct ICD-10 code: o A41.51 – Sepsis due to E. coli

4. The final code submission reflects the true complexity and acuity of the patient, without requiring the physician to change their natural language.

This isn’t science fiction. Natural language processing (NLP) and machine learning tools are already being used to draft clinical notes, predict disease, and flag sepsis in real- time. Applying the same logic to ICD coding would reduce unnecessary queries, ease physician burden, and allow hospitals to capture appropriate reimbursement based on existing documentation.

The benefits of this approach include :

• Reduced Physician Burnout: Physicians can document in the language they’re trained in, without needing to memorize ICD codes or worry about phrasing.  Improved Reimbursement Accuracy: Hospitals can ensure their DRGs match the true severity and complexity of their patients’ conditions.

• Fewer Queries and Interruptions: CDI specialists can focus on special cases instead of chasing down routine clarifications.

• Audit Readiness: Automatically derived codes based on structured data (labs, imaging, vitals) and NLP-reviewed notes can actually improve coding integrity. Companies like Empathia.ai, M*Modal, and Nuance are already building ambient listening tools to create clinical notes from conversations. The next logical step is to apply those same AI tools to derive the most specific ICD codes based on real-time structured and unstructured data in the EMR.

We must stop expecting physicians to become coders. Instead, let's build technology that understands us—and translates our clinical intuition into compliant, accurate, and financially sustainable documentation.

ICD coding and DRG payments are not going away. But it’s time we stopped letting them dictate how physicians document or how hospitals are paid. We have the tools to bridge the gap between clinical language and billing language. Let’s use AI to let physicians be physicians—and let machines handle the code translation.