Vatsala Muthukumaraswamy

Let’s ground this in my domain of inpatient medical coding, since knowledge management challenges are especially costly there. Knowledge Challenge (Relevance) One of the biggest wastes in inpatient coding comes from scattered reference materials: coders juggle ICD-10-CM/PCS codebooks, AHA Coding Clinics, MS-DRG logic, hospital-specific guidelines, and payer policies. Cross-check against the latest guidelines & facility rules. Highlight coding risks (e.g., “Possible MCC missed: Acute Kidney Injury per KDIGO criteria found in lab values”). Provide inline citations to the knowledge source → so coders trust the recommendation. This flow converts searching for knowledge into applying knowledge in real time.

Bias in AI is not a technical bug; it's a systemic threat that can cascade through healthcare outcomes, compliance, and trust. Bias Scenario: Automated Case Prioritization in Medical Coding Suppose an AI tool is employed to rank medical coding cases by urgency, complexity, or reimbursement value. If the training dataset disproportionately represents particular populations ( for example-older adults, inner-city hospitals, affluent zip codes), the AI will tend to rank those cases more often—discriminating against cases from rural, low-income, or minority communities. In this scenario, we should have medical coders from diverse backgrounds involved in the design process to highlight possible blind spots and enable them to alert suspicious prioritizations, inputting those instances back into the model for retraining.

This is a solid and comprehensive framework for AI solution handover and lifecycle management. AI solution purpose We should include the problem solved by AI, scope of work, data sources, KPI and ROI goals for shared understanding so that AI will not get misused or misinterpreted by anyone Technical To prevent “black box” syndrome, we should create architecture data flow diagrams, versions, data sources and update the schedule accordingly Functional User guides and video demos help to keep as learning assets Training Two days onboarding sessions which help reduce the dependency on the original stakeholders Performance KPI dashboard is useful to keep the performance tracking and continuous monitoring, and also to keep AI relevant and prevent silent degradation Change management SOP has to be updated to prevent accidental breakage and ensure traceability Escalation matrix Contact list of end-users, operational, and technical AI support to avoid downtime or any delay when issues arise Compliance Compliance checklist helps us to make sure legal and ethical longevity of the AI Roadmap To keep AI evolving and retain its value, future features have to be listed as proactive measure. This handover plan framework not only supports smooth handover but also embeds governance, accountability, and adaptability which are the pillars of sustainable AI.

The selected process is Reviewing inpatient medical coding quality Normally, a coding quality team undertakes the following responsibilities: Reviews a sample of coded charts Mistakes with flags, such as missing MCC/CC or using the improper DRGs gives coders input. Enhancements to the precision of accuracy scores using AI in several forms include for error prediction or DRG suggestion. Real-time scanning of coded charts anticipating possible coding errors or DRG mismatches is now possible by AI Spotting human-inspection, high-risk charts Pre-screening with AI: Every coded chart runs via AI, which notes possible issues (such as missed MCCs). Review by a Human-in-the-Loop: Reviewers or coders confirm flagged charts. Decisions are recorded as true or erroneous in AI Quality Analyst Track error trends, escalate patterns Train AI models using labeled examples, monitor drift, and identify false positives Team Leader assigns audits, generates reports, coordinates AI-human workflows, defines handoff points, resolves escalations. Align with coding standards, retrain ML models, and optimize prompt logic with the help of an AI Specialist / Clinical SME (who was previously not a member of the team). On a weekly or monthly basis, feedback is used to retrain models, improve prompts, or edit regulations. Lead + SME is responsible for mediating any conflicting coder input or low-confidence AI flag. Need New Talents This, though, is only successful if the process is modified to include AI handoffs. Using analytics and model governance improves roles. Openness and repetition help to develop trust.

In medical coding, a field that is constantly evolving, with rules, documentation practices, and payer policies changing, AI systems tend not to fail catastrophically. Rather, they tend to fail in a slower, more insidious way. These are the early indicators that AI systems are in decline Drop in MCC/CC Capture Rate (in relation to historical averages) – Suggests that the AI is missing important diagnoses which are clinically relevant Gain in Time-to-Code – Increasing time taken to code due to irrelevant suggestions and prompts Frequent No Suggestion Responses – The AI does not offer suggestions in cases where it used to provide them. Surveillance Focused on Prevention and Immediate Response We should establish baseline KPIs such as Pre AI Implementation and Post AI Implementation Take trends based on MDC, DRG, or MCC/CC capture rate Is there a more than 15% increase in time-to-code in three weeks against the historical average? That is a red flag Spot the missing idea streams (for example, a dip in suggestions for Pulmonary DRGs) Summary Track a combination of outcome signals (MCC/CC rates) and interaction signals (override feedback loops) Set defined automatic review boundaries Use simple real time drift trackers for monitoring Increase coder focus using AI output transparency and AI explainability.

In the realm of medical coding, where transparency, compliance, and alignment with business objectives are paramount, the audit framework known as "RAISE" is tailored for processes enhanced by artificial intelligence. The RAISE framework serves to assess the Reliability, Accountability, Impact, Security, and Ethics of AI systems within practical workflows. Reliability Is the AI system capable of consistently delivering reliable and accurate results, even in unusual circumstances? Accountability Is it possible to trace the origins of decisions, figuring out if they come from AI, a medical coder, or a system rule? Impact on Business and Compliance Does the AI provide results that comply with regulatory norms and fulfill ROI objectives? Security Are the components of the AI protected, allowing for regulated access and eliminating the possibility of misuse or unintentional data disclosure? Ethics Do AI systems make decisions that are fair, unbiased, and grounded in ethical principles? Checkpoints: We can determine if accuracy declines over time. We can maintain a log that lists responsibilities and specifies who is responsible for AI malfunctions. We can monitor regularly whether AI is compliant with the UHDDS standard, POA, and HAC regulations We can follow up and make sure who has the authority to see or alter AI models, prompts, or outcomes, as indicated in the access logs We can check whether the system explain the reason why it recommends a specific code, diagnosis, or DRG. Risk: The main risk is there is possibility of AI making high-impact decisions without any transparency. Thus, RAISE helps us to keep AI aligned with both patient care and business outcomes.

In medical coding, the AI systems are not “set-it-and-forget-it” tools. They should evolve with coding guidelines and updates. The signs that AI Solution in medical coding is becoming fragile The coders frequently ignore or override or bypass the AI suggestions The productivity time increases instead of decreasing The AI is not adapting to new or updated coding guidelines and appearing with old or outdated codes How to ensure long-term AI sustainability in medical coding Insert AI Governance Establish a cross-functional governance team and update the frequency such as quarterly validation against latest ICD-10-CM/PCS Plan Feedback loops Let coders capture corrections, override reasons, and user feedback data Build for Adaptability Use dynamic models and avoid brittle and hard-coded logic Monitor AI performance Set KPIs, monitor for drift, and validate outcomes regularly Tie to business value We should link AI to measurable impact and revalidate the business case every 6 months. Summary The implementation of sustainable AI in medical coding transcends mere technology. It involves integrating AI into an ever-evolving process. The leaders in Business Excellence should regard AI models as dynamic entities that necessitate governance, feedback, measurement, and adaptation.

In medical coding, the AI initiatives are based on value rather than mere novelty. The framework we use to keep AI efforts grounded in value is "Problem-First, Not Tool-First" The most fundamental and essential question to ask is "What ongoing business challenges are we addressing, and in what are the ways will solving these issues improve patient outcomes, compliance, or revenue?" How to apply in Medical Coding: We should always start with a Significant Problem. Find out the root causes of COPQ (Cost of Poor Quality) to uncover the pain points. For example, if there is any frequent errors on CC/MCCs which impact the case mix index. Quantify the Issue: What is the financial impact per 1000 charts? Is this related to coder variability, deficiency in clinical provider documentation, or audit results? Then only, we can Investigate the AI with the below questions: Can AI help us in identifying high-risk charts before the account sends for billing? Can NLP enable the automatic detection of missed or incorrect diagnoses or procedures? Does the solution reduce the manual effort, improve consistency, or prevent revenue loss? We should identify friction points in coding workflows by using VOC + COPQ + Pareto + Value Stream Mapping We should determine where AI could expedite and improve the processes, rather than merely automating for the sake of novelty. The transformation professionals can ensure real ROI and long-term adoption by consistently focusing AI discussions on measurable outcomes instead of merely on technological features.

In Swiss cheese model, the slices of cheese symbolize the defense layers such as processes, systems, and training which aim at preventing the errors. Holes in the cheese signify gaps in the defenses such as human errors, communication gaps and system constraints. Slices of Cheese (Defense Layers) Provider clinical documentation Defense: Educating the providers on best practices for documentation Software Encoder The tool that facilitates the assignment of appropriate codes Defense: Applying strong Poka Yoke technique in software encoder which prevents common errors from occurring Holes in the Cheese (Weaknesses) Poor Provider Clinical Documentation Missing specificity or type or acuity of diagnosis Not Hard Poka Yoke Control method in coding encoder Warning which is a soft Poka Yoke method in encoder that allows coders to accept unspecified and specified type of same diagnosis, and symptoms with definitive diagnosis How to use Business Excellence to Strengthen the System DMAIC Framework Example: Define Problem: High percentage of missed MCC/CCs which impact DRG reimbursement Goal: To Improve CC/MCC capture rate and reduce the rebill Measure The Baseline CC/MCC capture rate Analyze Find out which layers have the largest or most aligned "holes." For example - 80% of missed CCs linked to specific provider units Improve Strengthen defense layers: Implementing targeted provider education Incorporate first time right checklist within encoders Control Mistake-proofing: Introduce a "Swiss Cheese Risk Checklist" for coders Summary The Swiss Cheese Model assists inpatient coding teams in visualizing the interrelated risks and identify where defensive layers can be implemented or strengthened. By integrating it with Business Excellence tools such as DMAIC, KPIs, Voice of Employee, and Lean, we can transform from reactive correction to proactive prevention.

The change management is like a superglue that holds or grasps the DMAIC improvement together. How Change Management Fits Across DMAIC Phases Define We should build urgency and buy-in senior management, align all key stakeholders on the significance of the problem and need to improve MBB organized the Voice of Employee (VoE) sessions with coders and QA reviewers to find out pain points. These insights were used to articulate the problem statement: “We spend over 15 hours daily reworking charts that could be completed correctly the first time.” Measure We should communicate transparently about the current process gaps and baseline data MBB shared baseline audit data which revealed 18% rework rate with the top five error types. Analyze We should collaborate with cross-functional team and find out root causes MBB should ensure that the coders, and auditors should collaboratively identify the root causes at the time of fishbone and pareto analysis. The errors were outlined as process and training gaps (not coder incompetency), reducing resistance. Example: “If 65% of errors are due to DRG mismatch, what process steps enable this?” Improve We should solve “what’s in it for me” (WIIFM) for all impacted groups, find out the solutions, pilot and celebrate quick wins Instead of giving solutions, MBB facilitated sessions for the team to share their ideas. Coders proposed a new DRG cross-check checklists and mandatory peer review for new coders. These initiatives were piloted within one team for four weeks, which resulted in 40% reduction in rework. The quick wins were shared in townhalls to keep the momentum active. Control Sustain the solutions by creating new SOPs, visual management, dashboards, and feedback loops. MBB established a visual management dashboard which displayed monthly rework rates by team, integrated weekly error review huddles, and nominated a Coding Quality Champion from the team. They Conducted 30-60-90 day audits following the project to confirm sustained improvements. The best-performing coders were publicly recognized to maintain motivation. Project Outcomes: Rework rate decreased from 18% to 9.5% within 5 months Average Turnaround time improved by 10 hours Client audit satisfaction score increased from 91% to 96% Coders reported higher confidence and lower frustration (as indicated by the VoE survey) Role of MBB in Driving Adoption: VoE engagement and listening sessions Collaborative identification of root cause analysis and solutions Celebrating quick wins Integrating visual control systems Institutionalizing new practices through SOP revisions and dashboards Ensuring accountability handover post-project In this project, the technical solutions such as checklists and peer reviews were significant. However, it was the coders’ emotional engagement, empowerment, and shared ownership facilitated through disciplined change management that made the improvement stick. That represents the true value added by MBB beyond Six Sigma tools.

Even after a process improvement yields significant benefits such as reduced coding errors, and increased coding accuracy, if the Control phase is not robust, these advancements can easily diminish. Reasons for the Decline: - Lack of dashboard or KPI tracking following implementation, leading to a gradual increase in unnoticed errors - New coders may not be aware of the latest DRG validation checks or coding clarifications - The Initial excitement and enthusiasm fade away, and without any accountability, the old habits appear again - The newly released CC/MCC guidelines are not integrated into coder checklist - There is no recognition for coders who consistently sustain progress Tools and Techniques: - Utilize a control chart to monitor weekly DRG change rates after PCS coding enhancements - Revise your CC/MCC capture checklist to include newly added conditions from last improvement cycle - A Leaderboard of Coding Accuracy Champions or Coding Ninjas which can be visible on Teams/SharePoint - A pop-up warning if a major OR procedure is missing in a high-risk DRG chart. Control plan document: - Develop a control plan document that outlines what will be monitored, who will be responsible for tracking, the frequency of tracking, target thresholds, and action plans for any outliers.

If medical coders or auditors recognize any persistent issues, or if Business Excellence teams alter workflows (for instance, by modifying DRG validation checks or POA flagging protocols), those insights are frequently confined to process documentation, unless they are explicitly reintegrated into AI retraining or rule modifications. How to Ensure AI Evolves with the Process in Medical Coding - Practicality of Integrating AI into Improvement Cycles This is achievable only if AI is regarded as an ongoing participant in the process instead of a fixed tool. In Medical Coding, regular coding audits, patterns of errors, and feedback loops from inquiries can be systematically gathered. Feedback Mechanism utilize organized, documented, and classified AI error logs, accompanied by established retraining timelines and configuration modifications. MBB Role Act as stewards for AI process alignment, oversee AI feedback cycles, spearhead AI-impact PDSA/DMAIC initiatives, and promote governance for enhancements in AI.

In Medical coding, AI solutions are usually applied to automate the coding process, reduce errors or enhance revenue cycle management. If problem is not clearly defined, even technically proficient AI may not deliver value, because it may focus on symptoms instead of root causes. Master Black Belts (MBBs), with their knowledge of Lean Six Sigma methodologies, process enhancement, and data-driven problem-solving, are ideally positioned to ensure that the correct problem is identified. Example: AI in Medical Coding Addressing the Incorrect Issue Scenario: A hospital aims to decrease claim denials within its medical coding process, blaming the problem on coder mistakes in assigning ICD-10 codes. An AI system is created to automate code assignments using historical patient data, achieving a 95% accuracy rate in code prediction. However, claim denials continue to be high, and coders express frustration with the AI overriding their expertise in complicated cases. Why It Fails: The issue was misidentified as "coder errors in ICD-10 assignment" when the actual cause is inadequate clinical documentation (for instance, vague or incomplete physician notes). The AI, trained on existing records, accurately assigns codes based on the data available but cannot rectify documentation deficiencies, which lead to denials. The issue here is misinterpreted as "coder errors" instead of "poor clinical documentation". Hence, the claim denials are not reduced by AI effectively. When AI tackles a wrong issue, the resources are not able to provide any impact and there is dissatisfaction among stakeholders. This can be prevented by Master Black Belts by implementing DMAIC, conducting root cause analysis, engaging key stakeholders, validating assumptions, ensuring data correctly, and integrating change management. If the problem is reframed to focus on the quality of documentation, MBBs confirm AI solutions can give meaningful outcomes, such as reduction in denials and improved efficiency in the revenue cycle.

In medical coding, it is possible for AI to bring transformation in effectiveness and efficiency. But this should be governed with precision and accuracy. We should define the permissible usage, constraints, and ethical guidelines for AI applications in programming and ensure compliant with HIPAA standards whenever applicable. Medical Coding Specialists validate AI-generated codes and offer continuous training feedback mechanisms. The real-time dashboards like Kanban boards could be executed to visualize and monitor the coding accuracy With respect to Explainability & Traceability, every AI-generated code is auditable and linked to specific medical record documentation. The outputs are integrated explaining "why this code was selected" as part of the coding workflow and facilitating transparency. The corrections are made to retrain models and ensure continuous enhancement is proper aligned with Business Excellence. We can start with less critical coding areas like outpatient coding with pilot sandboxes To control mechanisms, role-based access is to be given to sensitive information like protected health information. There should be governance board reviews for significant model updates or any changes in rule logic. There should be transparency in terms of clarifying coding logic, sustaining audit trails, and involve key stakeholders in decision-making activities.

Yes, ready. There exists a relationship between Business Excellence (BE), AI transformation, and the measurement of organizational performance. Traditional metrics of Business Excellence were effective in systems driven by human input and repetitive transactions. AI is now responsible for decision-making, process optimization, and large-scale output personalization, some of these metrics may become misleading, incomplete, or even outdated. In medical coding field, the existing metric which is possible to be outdated/obsolete is the number of charts coded by coder in an hour. Earlier, the medical coder was able to code 25 charts per hour, but now they may only code 15 charts per hour because of complexity and AI codes 30 charts which are all easy and simple. Subsequently, the obsolete productivity metrics will show a decrease in trend. The new metric to introduce is Human-AI Collaboration Ratio (HACR) in Coding. It tracks how AI augments and not replaces workforce capacity The updated/revised business excellence scorecard for AI-enabled medical coding is for Productivity - The traditional metric is the number of charts coded per FTE. The updated AI-enabled metric is the total number of charts processed (AI + human) and HACR. How it is related to Business Excellence: These new metrics: • Connect AI performance with human empowerment • Assist leaders in proactively managing AI drift and coder disengagement How to Implement: • Test new metrics alongside existing ones to ensure continuity and facilitate comparison • Utilize heatmaps to illustrate performance by specialty or AI version, which is beneficial for CAC tuning • Ensure dashboards are aligned with compliance, revenue integrity, and training functions, not solely operations.