Medical Coding Accuracy

Why Medical Coding Is So Hard to Automate Last week, I talked about the challenge of building an autonomous medical coding tool. The biggest hurdle? Generative AI’s unpredictability. But beyond that, there’s a deeper issue: medical coding isn’t just about knowing the codes—it’s about knowing how to apply them correctly, which is highly payer-dependent. A basic coding job? Sure, you can Google ICD-10 codes or use online lookups. But when it comes to real-world coding—determining the principal diagnosis, applying correct bundling modifiers, and ensuring compliance with payer-specific policies—the complexity multiplies. Take orthopedics as an example: - If a provider performs an evaluation and an injection in the same visit, modifier 25 is required to ensure proper reimbursement. - But different payers have different interpretations of when this applies, meaning the same case could be coded differently based on who is paying the bill. And that’s just one modifier. Every year, coding policies evolve, fee schedules change, and payers tweak their criteria for approving or denying claims. Some of these updates happen annually, but others roll out throughout the year, requiring constant monitoring. That’s why a true automation solution for medical coding can’t just be “smart.” It needs to be adaptable. The right tool must: ✅ Continuously update with new CPT codes, bundling rules, and compliance requirements. ✅ Understand payer-specific nuances, instead of just applying generic coding rules. ✅ Reduce the mental math that coders and billers do daily to navigate complex documentation and reimbursement policies. Without this level of customization, coding automation just shifts the burden back to staff. What’s your biggest frustration with medical coding today? Have payer policies made it harder to get claims through?

Manual ED ICD-10 coding is error-prone and slow, so the authors test a retrieval-augmented generation (RAG) pipeline that reranks candidate codes from a 1,038,066-visit database to help LLMs assign more accurate and specific primary diagnoses. 1️⃣ In a retrospective case study of 500 ED discharges from six Mount Sinai hospitals (Jan 1–Apr 30, 2024), nine LLMs generated zero-shot codes and then RAG-enhanced codes restricted to one of the top-10 retrieved ICD-10-CM descriptions. 2️⃣ Valid-code yield improved markedly with RAG (for example, Qwen2-7B from 43.4% to 99.8%; GPT-4 from 95% to near-perfect), and exact human–LLM code matches also rose (GPT-4 from 36 to 132; Qwen2-7B from 4 to 88). 3️⃣ Masked human review favored RAG-enhanced GPT-4 over provider codes for accuracy in 447 vs 277 instances and for specificity in 509 vs 181 (both P<0.001), while LLM reviewers showed similar preferences and moderate-to-high interrater agreement patterns. 4️⃣ The paper recommends phased deployment with EHR integration, clinician-in-the-loop review, prospective A/B testing, continual updating, and local RAG databases to address generalizability, transparency, and overcoding risk. ✍🏻 Eyal Klang, Idit Tessler, Donald Apakama, Ethan Abbott, Ben Glicksberg, Monique Arnold, MD, Akini Moses, Ankit Sakhuja, MBBS MS FACP FASN FCCP FCCM FAMIA, Ali Soroush, MD, MS, Alexander Charney, David Reich, Jolion McGreevy, Nicholas Gavin, Brendan G. Carr, MD, MA, MS, Robert Freeman, Girish Nadkarni. Assessing Retrieval-Augmented Large Language Models for Medical Coding. NEJM AI. 2025. DOI: 10.1056/AIcs2401161

Step 4 of the Revenue Cycle Management (RCM) process is here, and this is the part where accuracy becomes money. Coding & Charge Entry. The step that quietly decides whether your claims get paid… or denied. Let’s break it down in the simplest way possible. What Is Coding & Charge Entry? Coding is the process of converting provider documentation into standardized medical codes: - CPT - ICD-10 - HCPCS Charge Entry is the next step, entering those codes into the billing system with the correct fees, modifiers, and patient information. Think of it like translating a provider’s work into a language insurance understands. If the translation is wrong… the payment won’t come. Why This Step Is So Important Accurate coding leads to: ✅Clean claims ✅Faster reimbursements ✅ Fewer denials ✅Strong compliance ✅Accurate reporting ✅Transparent patient billing One wrong code can delay payment for weeks. One missing modifier can trigger an instant denial. One documentation gap can cause revenue leakage. This is the step where precision pays , literally. 📌 Example A patient comes in for a simple wound repair. The provider documents: - 2.5 cm laceration - Simple closure - Local anesthesia Correct code: 12001 But if someone mistakenly enters 12002 (intermediate repair), insurance will deny it. Not because the work wasn’t done… But because the code doesn’t match the documentation. This is why Coding & Charge Entry must be clean, accurate, and detail-driven. 🛠️ How a Virtual Medical Assistant Helps A skilled VMA supports providers by: ✅Reviewing documentation ✅Assigning accurate codes ✅Checking payer-specific rules ✅Ensuring correct modifiers ✅Preparing clean charges ✅Flagging missing information ✅Coordinating with billers and providers Result: 📌Fewer denials. 📌Faster payments. 📌Better patient experience. 📌Stronger revenue flow. If Step 3 is about capturing what happened, Step 4 is about converting it into revenue. And in a busy practice, accuracy is everything. If you’d like help improving your coding and charge entry workflow, I’m here to support your team with reliable, detail-driven assistance.

The Impact of Accurate Documentation: Real-World Examples from an OIG Audit The importance of accurate condition capture and documentation in risk adjustment cannot be overstated. A recent OIG random review of EmblemHealth highlights just how much is at stake when conditions are incorrectly reported—or overlooked entirely. Here are three patient scenarios uncovered in the audit: Enrollee A • Submitted HCCs: 3 • Validated HCCs: 2 • Review Findings: The submitted HCC for Polyneuropathy was not supported by the medical record. • Impact: $1,992 overpayment to the plan. Enrollee B • Submitted HCCs: 2 • Validated HCCs: 2 • Review Findings: While the submitted HCC for Diabetes Mellitus (DM) with complications was not supported, the documentation did validate DM without complications. • Impact: $2,328 overpayment to the plan. Enrollee C • Submitted HCCs: 2 • Validated HCCs: 4 • Review Findings: Reviewers not only validated the submitted HCCs but also identified two additional HCCs that had not been submitted in claims. • Impact: $4,438 underpayment to the plan. The Takeaways The OIG review illustrates several key themes: 1. Documentation Gaps: Unsupported conditions can lead to overpayments, exposing plans to potential financial penalties and compliance risks. 2. Missed Opportunities: Inaccurate or incomplete submissions may result in underpayments, reducing the resources available to care for complex patients. 3. Importance of Proactive Reviews: Comprehensive reviews of documentation and coding practices can identify discrepancies before they escalate into larger issues. A Path Forward: Best Practices To mitigate risks and ensure accuracy: • Train Providers Continuously: Reinforce the importance of detailed documentation that supports all submitted diagnoses. • Conduct Pre-Bill Reviews: High-risk HCCs should undergo additional scrutiny before claims submission. • Leverage Retrospective Audits: These reviews can identify and address both unsupported claims and missed conditions, ensuring accurate risk scoring. The financial and compliance stakes are high. The EmblemHealth review revealed a $552,000 overpayment recommendation from the OIG, with 25% of submitted conditions unsupported by medical records. On the flip side, 65 additional conditions were identified but not submitted, representing missed opportunities for accurate risk adjustment. In risk adjustment, accuracy is everything. How does your organization close the gap between documentation and coding? Let’s connect and share strategies.

#TipTuesday: Why You Should Avoid Unspecified Diagnosis Codes Unspecified diagnosis codes (those ending in ".9" or "0") can be tempting when documentation lacks detail, but they can trigger denials, payment delays, and compliance risks. Here’s why you should steer clear of them and what you can do instead: Why Unspecified Codes Are Problematic: ❌ Payer Scrutiny – Insurance companies may flag unspecified codes as insufficient for medical necessity. Some outright deny claims with them. ❌ Delayed Reimbursement – Payers may request additional documentation, slowing down payments. ❌ Higher Audit Risk – Frequent use of unspecified codes could suggest poor documentation practices and lead to compliance issues. ❌ Lost Revenue – Some payers downcode unspecified diagnoses, leading to lower reimbursement rates. When Are Unspecified Codes Acceptable? 🔹 New Patient or Initial Visit – If a provider is still assessing a condition, a temporary unspecified code may be justifiable. 🔹 Limited Documentation – If a provider lacks sufficient details (e.g., ER settings), unspecified codes might be necessary. 🔹 Payer Guidelines Allow It – Some insurers accept them in certain circumstances, but always verify. How to Reduce Unspecified Code Usage: ✔ Educate Providers – Train clinicians on the importance of detailed documentation. Encourage them to specify laterality, acuity, and underlying conditions. ✔ Query for Clarification – If the documentation is vague, don’t guess! query the provider for more details. ✔ Utilize Coding Resources – Use encoders, coding books, and payer policies to ensure you’re selecting the most specific code. ✔ Track & Audit Coding Trends – Identify patterns in unspecified code use and implement corrective actions. Specificity is key to accurate billing, clean claims, and compliant coding. Avoid the temptation of unspecified codes. Your revenue cycle will thank you! #TipTuesday #MedicalCoding #MedicalBilling #RevenueCycle #HealthcareCompliance #CoderLife

One coding mistake I see all the time and it’s costing accuracy and revenue. Next time you see 'contusion' in a trauma chart… what will you do differently? 📌 Basics of Reporting Injuries – Know the Order of Severity When multiple injuries happen at the same site, always report based on severity. Here’s a simple reference list 👇 🔹 Fracture – Bone is broken 🔹 Dislocation – Joint is out of place 🔹 Sprain – Ligament injury 🔹 Strain – Muscle/tendon injury 🔹 Laceration – Deep cut or tear 🔹 Contusion – Bruise 🔹 Abrasion – Surface-level scrape ✅ Always code the most severe injury first (PDX😉) ✅ Code injuries separately if they occur on different body parts 🚨 A common mistake I see: Misreporting Contusion/Laceration. Why? Because we usually think of contusion as a minor skin-level bruise. That’s true in musculoskeletal cases. But what if it’s an internal organ contusion or laceration?🤔 For example, a splenic contusion is a sign of internal organ damage. This is much more severe and should never be coded like a skin bruise. Real-life case scenario: A patient presents after a road traffic accident with: 🎯Traumatic cerebral hemorrhage 🎯Splenic laceration 🎯Rib fractures 🎯Multiple abrasions on limbs Which one takes priority? 🩺 The internal injuries – cerebral hemorrhage and splenic laceration. Doctors will always stabilize the life-threatening conditions first, like brain bleeds or internal bleeding. Getting the sequencing right isn’t just about coding , it’s about reflecting the clinical reality and ensuring accurate reimbursement. As a coder or CDI, your documentation and code assignment should reflect that priority. ✅ Always query if severity is unclear ✅ Check for terms like hemoperitoneum, shock, or emergency procedures like craniotomy or splenorrhaphy ✅ These affect both clinical understanding and reimbursement So next time you see “contusion,” ask yourself: 📌Is it a simple bruise, or is it internal organ damage? Your answer can change the entire severity and weightage of the case.😄 If you're a new budding coder and found this helpful, follow along! If you're an experienced pro😎, drop your personal insights in the comments below. Because let’s be honest 🤨 we all felt lost when we first started coding without any background in these terms, right? #CDI #ClinicalDocumentation #MedicalCoding #InjurySeverity #Reimbursement #HealthcareLearning #CodingTips 

Let me let you in on something auditors see every single day but rarely explain to physicians. “Unspecified.” You write it when you are not 100% sure. It feels safe. It feels honest. It feels like the responsible thing to do when you do not want to overstate. But here is what nobody tells you. To a coder and a payer, unspecified does not read as cautious. It reads as low severity, low acuity, low reimbursement. You think you are playing it safe. What you are actually doing is erasing how sick your patient really is. And it happens on claim after claim after claim. Let me show you exactly what I mean. Heart failure unspecified versus acute on chronic systolic heart failure. Same patient. Two completely different stories — and two completely different reimbursements. Anemia unspecified versus acute blood loss anemia after surgery. One quietly downgrades the claim. The other reflects what actually happened in that OR. Stage 3 CKD documented on its own versus Stage 3 CKD linked to hypertensive heart disease. That one link changes the entire picture. Now here is the part I really need physicians to hear. Specificity is not upcoding. It is accuracy. Documenting the true severity of the patient sitting in front of you is not gaming anything — it is the system working exactly the way it was built to. The diagnosis is not just a label you drop in the chart. It is the story of how sick your patient is. And unspecified erases that story before a coder ever gets to tell it. After 25 years of auditing this is hands down the easiest fix I can hand a physician. Be specific. Acute or chronic. Right or left. Linked or unlinked. Every single time. So be honest with yourself — are you documenting the diagnosis, or just defaulting to unspecified and hoping the coder figures out the rest? Drop your thoughts below. #CodeLikeAnAuditor #ClinicalDocumentation #CDI #MedicalAuditing #HCCCoding #PhysicianDocumentation #QueensCodeAI

🚨 Most “HCC coding” tools just chase more diagnosis codes. In value-based care, real risk adjustment needs much more. 🚨 Accurate HCCs are context- and vocabulary-aware. They don't just keyword match; they respect CMS rules and keep the clinical story intact instead of blindly inflating RAF scores. What should medical LLMs for HCC coding actually do? • Synthesize the full chart: Connect problems, meds, labs, and imaging, not just isolated notes. • Orchestrate the vocabulary stack: Validate ICD-10-CM diagnoses against LOINC labs, NDC medications, and CPT/HCPCS procedures to prove the "MEAT." • Flag the unsupported: Identify conflicting or outdated conditions rather than auto-accepting every suggestion. • Ditch static memorization for live retrieval: Codes change. Formularies update. Models must query live ICD, CPT, and NDC datasets in milliseconds, not rely on stale training data. • Provide an audit trail: Show the exact path: Phrase → Concept → Code → HCC (with confidence scores). How many open, production-ready stacks do all of this today? Not many. That gap slows honest benchmarking and safe automation. What if HCC capture were transparent with LLMs + live code sets working together, instead of a black box? Should OpenMed ship an open, audited, policy-aware HCC coding stack with live vocabulary retrieval? If that would help your organization, drop a comment or DM me. Let’s democratize HCC coding and make risk adjustment decisions transparent.

🔔 The HHS Office of Inspector General just released its first industry-specific Compliance Program Guidance for #MedicareAdvantage. This isn't routine. This is the government telling the industry: the rules of engagement have changed. The core question the MA ICPG poses to every health plan is simple. Can you prove that what you submitted is accurate, clinically supported, and tied to real patient care? Not just coded. Cared for. For years, risk adjustment has operated as a revenue function. Find more codes. Increase RAF scores. Optimize reimbursement. The MA ICPG makes clear that this approach now carries real regulatory and legal exposure. The new standard is defensible coding — every diagnosis encounter-linked, clinically evidenced, and audit-ready. The guidance directly addresses the patterns that have drawn federal scrutiny. Retrospective programs that only add codes and never remove unsupported ones. AI tools prompting providers to document conditions without clinical basis. Assessments generating diagnoses with no follow-up care. Going forward, the OIG expects: → Your risk adjustment vendors are now a direct compliance obligation → Your retrospective programs must detect ADDs & DELETEs → Your AI must be explainable, auditable, and governable ⚠️ Every health plan leader should read this guidance not as a regulatory update but as an operational stress test. Does your current program hold up? Do your vendors? Does your technology? At RAAPID INC, we built our platform around this exact standard. Our Neuro-Symbolic AI delivers defensible coding with transparent evidence trails for every recommendation. Two-way retrospective coding. Encounter-linked documentation. Every diagnosis is grounded in clinical truth — not just coding logic. Our Neuro-Symbolic AI doesn't just find codes. It explains why, with proof. Everyone in risk adjustment knows the old model was on borrowed time. The OIG just put a date on it. ✅ Defensible Accuracy. Proactive Compliance. #RiskAdjustment #DefensibleCoding #NeuroSymbolicAI

Day 54/365 - #MedicalCodingInsights 🔍 Mastering the Steps in the Medical Coding Process (ICD-10-CM) 🧾 Accurate medical coding isn’t just about assigning numbers — it’s about following a systematic, validated process to ensure compliance, reimbursement, and quality reporting. Here’s a simplified breakdown every aspiring coder must master: ✅ 1. Identify the Main Term Always start with the core condition, not the modifier. Example: Chronic renal failure → Main term is Failure, not chronic or renal. ✅ 2. Search the Alphabetic Index Follow this exact order: 🔹 Condition → Organ/Anatomical site → Manifestation → Modifiers → Laterality ✅ 3. Follow Cross-References & Notes Never ignore instructions like “See,” “See also,” or category notes — they guide you to the most accurate code. ✅ 4. Apply Modifiers Correctly Understand the difference between: 🔹 Essential modifiers (affect code selection) 🔹 Nonessential modifiers (clarify but don’t change the code) ✅ 5. Verify in the Tabular List The Tabular List is the final authority — always confirm laterality, instructional notes, and character requirements. ✅ 6. Code to the Highest Specificity Use all required 4th–7th characters, placeholders (X), and encounter extensions (A, D, S) when applicable. 💡 Expert Coding Tips: ✔️ Read the diagnosis right-to-left if confused ✔️ Avoid unspecified codes when laterality exists ✔️ Always confirm medical necessity ✔️ Never code directly from the Alphabetic Index alone 📘 This structured approach ensures clean claims, fewer denials, and strong compliance when working with ICD-10-CM. If you’re building a career in Medical Coding, Billing, or AR, mastering these steps is non-negotiable 🚀 👉 Save this post for revision & share with fellow coders! 🔖Follow Kartik Khedkar to get more insightful updates! #MedicalCoding #ICD10CM #HealthcareDocumentation #RCM #MedicalBiller #CodingLife #HealthcareCareers #ClaimAccuracy #MedicalAudit