Clinical Documentation Improvement

Most MedTech companies treat audits as one-off events. (And it costs a lot more than money) This mindset costs: • Market access • Investor trust • Years of work product • And lots of money    But the biggest cost isn't financial. It's human lives. The ones that depend on life-saving devices that are getting locked out of the market. Not because their technology wasn’t good enough. But because of preventable mistakes. Because they treated compliance as an event. Not a culture. Passing a Notified Body Audit isn’t luck. It’s discipline. It’s daily habits. It’s system-level thinking. Here are 4 ways the best MedTech companies prepare (and how you can too): 1. They build audit-ready systems Your documentation must tell a complete story: • Align QMS to ISO 13485:2016 and MDR Article 10 • Justify risk management with defensible rationales • Show proactive surveillance in PMS reports • Close CAPAs fully with evidence of resolution • Validate claims with clinical performance data 2. They eliminate silent compliance risks Fix problems that quietly undermine audits: • Complete missing risk–benefit rationales • Update and control all key documents • Close gaps in complaint and vigilance logs • Strengthen post-market surveillance • Link CAPAs directly to audit findings 3. They train for audit readiness every day. Turn audit behavior into muscle memory: • Run mock audits and rotate team roles • Train clear, non-speculative auditor responses • Assign scope ownership across all functions • Focus answers — no speculation or improvisation    4. They set up audit execution in advance. Plan logistics that create calm, not chaos: • Prepare a dedicated audit room with indexed files • Assign document fetchers and tech support • Track requests and responses live during audits • Maintain a calm, professional audit environment Here’s the truth: An audit isn’t something you survive. It’s a mirror that reflects how you operate every day. What’s the biggest audit challenge your team is facing right now? ♻️ Find this valuable? Repost for your network. 💡 Follow Bastian Krapinger-Ruether for actionable tips on MedTech compliance and QM.

Healthcare providers, here are my top 5 documentation tips (coming from an RN and chart reviewer!) 1. Be Specific With Diagnoses 🔹 Instead of: "Diabetes" Write: "Type 2 diabetes mellitus with chronic kidney disease, stage 3" Why it matters: Specificity supports risk adjustment, accurate coding (HCC), and better treatment planning 2. Close the Loop on Abnormal Findings 🔹 Instead of: "Abnormal ECG – follow-up pending" Write: "ECG showed LVH. Will refer to cardiology and repeat in 6 months" Why it matters: Shows clear clinical reasoning and avoids appearing negligent in follow-up care. 3. Tie Medications to Diagnoses 🔹 Instead of: Just listing meds Write: “Patient on metoprolol for atrial fibrillation and HTN” Why it matters: Confirms the diagnosis is being treated and helps justify prescriptions and coding. 4. Reconcile and Update the Problem List 🔹 Remove resolved problems or mark them as inactive Why it matters: Keeps the chart clean, reduces confusion during transitions of care, and supports accurate billing. 5. Avoid Copy-Paste Without Updating 🔹 If using a template or previous note, always edit Why it matters: Reduces risk of errors, prevents contradictions, and reflects accurate clinical thought. These small tweaks can make a big difference in how your notes are interpreted, coded, and used by the rest of the care team. What would you add to this list? Let’s keep learning from each other.

Clinical trial analytics is messy… especially when you’re stuck between raw data ingestion and the actual analysis layer. I spoke with Mike Araujo from Medidata Solutions (life sciences tech company, ~25 years in the game) about their Data Connect and Clinical Data Studio solutions, designed to bridge that gap and put analytical power back in customers’ hands. They’ve built on AWS with: Apache Iceberg as the table format, hosted in the AWS Glue Data Catalog  AWS Glue and connected query engines for analytics – https://lnkd.in/dcXcyxBJ Amazon Managed Flink and managed Kafka for streaming – https://lnkd.in/ecxin49N This combo lets customers: Build and manage custom datasets themselves (they’ve already created 4,000+ inside the platform) Work with data via code (SDKs), UI-based interfaces, and soon, AI agents for users who know what they want but don’t have the technical details Plug their existing tools into the same Iceberg/Glue backbone without constant re-architecting Mike called the Iceberg + Glue Catalog setup “the straw that stirs the drink”, and AWS helps with the painful parts of running Iceberg at scale (like compaction and orphan file cleanup) so teams can “set it and forget it.” If you’re just starting: Take a hard look at where your data lives today, then figure out how to move toward a Glue + Iceberg backbone. Once that’s in place, a lot of the rest “falls into place.”

Hate how boring and time-consuming documentation feels? Yeah, same. But here’s the thing: the more you avoid it, the more you hurt your future self and miss opportunities to showcase your skills properly. So if you want to make documentation less painful (and actually useful), here are 6 tips I use with my clients to make it faster, clearer, and more impactful: 1. Start with an overview What’s the purpose of your project? What problem did it solve? Just 3–4 lines to set the stage. Make it easy for anyone to understand why it matters. 2. Walk through your process Break down the steps: How did you collect the data? How did you clean, analyze, or model it? What tools or methods did you use? This shows how you think and how you solve real-world problems. 3. Add visuals A clean chart > a wall of text. Use graphs, screenshots, and diagrams to bring your work to life. (And bonus: you’ll understand it faster when you come back later.) 4. Show your problem-solving What roadblocks did you hit? How did you fix them? Don’t hide your struggles, highlight them. This is where your value really shines. 5. Summarize your results What did you find? Why does it matter? What’s next? Answer these three questions clearly and your audience will instantly get the impact of your work. 6.  Use a structure that makes sense Try this flow: Introduction → Objectives → Methods → Results → Conclusion → Future Work Simple. Clean. Effective. P.S: After every milestone, take 5 minutes to update your notes, screenshots, or results. Turn it into a habit. ➕ Follow Jaret André for more data job search, and portfolio tips 🔔 Hit the bell icon to get strategies that actually move the needle.

Most AI audit programs test the wrong things. 😓 Not because the people running them don't know what they're doing, but because AI systems require three fundamentally different types of audits, and most programs are only running one. The core problem with applying traditional audit frameworks to AI is that traditional audits operate in a world of deterministic systems. A financial audit verifies whether transactions were recorded correctly. An IT audit verifies whether access controls functioned as designed. There is a correct answer, and you're checking whether reality matches it. AI systems are often probabilistic. A model never achieves perfect accuracy. That requires a different auditing logic entirely. Governing Intelligence by Noah M. Kenney defines three overlapping audit types, each asking a fundamentally different question: 1️⃣ Technical Audit: "Does this system work as designed?" This is the audit most teams are running. It covers model accuracy on holdout test data, performance consistency across demographic subgroups, robustness against adversarial inputs and distribution shift, and edge case behavior. Test sets must be representative of real deployment conditions, not just training conditions. A model tested only on data similar to its training set will look far more capable than it is. Subgroup testing is non-negotiable. 2️⃣ Algorithmic Audit: "Is this system fair, and does it reflect the values we've stated?" This is the audit most teams are not running systematically. It requires defining fairness metrics appropriate to the decision context, measuring whether the model meets those metrics, and acknowledging that no single fairness definition is universally correct. Demographic parity (equal outcome rates across groups), equalized odds (equal error rates across groups), and calibration (equal accuracy of predictions across groups) cannot all be simultaneously satisfied when true outcome rates differ across groups. This is the Impossibility Theorem in practice. The governance obligation isn't to satisfy all fairness metrics, it's to consciously choose which metric applies to your context, document why, and accept accountability for that choice. 3️⃣ Compliance Audit: "Does this system meet the regulatory requirements that apply to it?" This is documentation review, process verification, and regulatory gap analysis. Does the system have a completed Data Protection Impact Assessment? Is technical documentation current and accurate? Were conformity assessment requirements met before deployment? Are incident reporting procedures in place and tested? Are human oversight mechanisms functional rather than ceremonial? The compliance audit catches the gap between what governance documents claim and what governance infrastructure actually exists. Drop a comment on which of the three audit types is most underdeveloped in your program right now? #AIGovernance #AIAudit #GRC #RiskManagement #Compliance

How to Conduct Effective Internal Audits: Internal audits aren’t just a compliance requirement they’re a strategic tool for continuous improvement, risk reduction, and operational excellence. To ensure your internal audits deliver real value, follow these three critical phases: ✅ 1. Pre-Audit (Planning & Preparation): ◾ Define the scope, objectives, and criteria ◾ Review past audits and key documents ◾ Develop audit checklists and communicate with auditees ⚙️ 2. In-Process Audit (Execution): ◾ Conduct an opening meeting ◾ Observe processes, review records, and interview staff ◾ Document findings objectively (non-conformities, observations, OFIs) 📋 3. Post-Audit (Reporting & Follow-up): ◾ Hold a closing meeting to present findings ◾ Prepare a clear, concise audit report ◾ Ensure timely corrective and preventive actions (CAPA) ◾ Follow up to verify effectiveness Pro Tips: ◾ Maintain objectivity and confidentiality ◾ Use technology to streamline documentation and tracking ◾ Foster a culture where audits are seen as a tool for growth, not inspection ◾ Effective internal audits help organizations stay compliant, improve processes, and build a culture of accountability and excellence. #InternalAudit #AuditExcellence #QualityManagement #Compliance #RiskManagement #ContinuousImprovement #ISO9001 #AuditorLife #ProcessImprovement #HSE #InternalControls #OperationalExcellence #QHSE #Leadership #CorporateGovernance #CAPA #AuditTrail #ProfessionalDevelopment

✅ Fractional Improvement: Document the Doing You know that task you do every month that takes 45 minutes, even though it should only take 15? Yeah, that one. This week’s improvement is about turning that “I’ll just do it quickly myself” into something scaleable: 📹 Recording Looms while I work, not after 📁 Capturing process steps live in Notion 🧠 Creating a “How To” hub so the team (or future me) can take it on Why now? Because repeating yourself isn’t leadership. And efficiency isn’t about speed - it’s about transferability. My new rule: if I do it more than twice, it gets documented. No big SOPs. Just enough to make handover easy and errors unlikely. ✨ Fractional Improvement ✨ This one’s for the future-you who’s tired of refiguring it out every time. How do you document while doing, without slowing down? #FractionalImprovement #ScalingSmart #ProcessNotPain #FractionalOperations

It's #AI Friday! How does an algorithm assign codes in an autonomous medical coding solution? Imagine a smart robot that's learning to understand doctors' notes. First, it cleans up the notes to make them easy to read. Then, it looks for key words that tell it about the patient's health, like "cough" or "fever". It's like the robot is in school, studying lots of examples to learn how to match these words with medical codes, which are like secret codes that summarize the patient's condition. Once it's learned enough, the robot tries to guess the right codes on its own when it sees new notes. Sometimes, it checks its guesses with a teacher or gets better by learning from its mistakes. Over time, this robot gets smarter and better at its job, just like a student becoming an expert. Machine learning algorithms assign codes through a process that involves several key steps --- 💽 Data Preprocessing: The algorithm begins by preprocessing the input data, which typically includes clinical documents such as physician notes, discharge summaries, and other medical records. This step involves cleaning the data, normalizing text (like converting all text to lowercase), and possibly converting the text into a format that the algorithm can process more effectively, such as tokenizing sentences into individual words or phrases. ⚕️ Feature Extraction: The algorithm then extracts features from the preprocessed text. Features could include specific medical terms, phrases, contextual clues, and other relevant information from the text. Advanced machine learning models, especially those based on deep learning, can automatically learn to identify relevant features from the data. 👩🏫 Model Training: A machine learning model is trained on a dataset that has been manually coded by human experts. This involves feeding the extracted features and the corresponding codes into the model, allowing it to learn the complex relationships and patterns between the text features and the medical codes. This step may involve supervised learning, where the model is explicitly taught the correct codes for given features. ⏩ Coding Prediction: Once trained, the model can then predict medical codes for new, unseen clinical documents. It does this by extracting features from the new documents, similar to how it was trained, and then using the learned patterns to predict the most likely codes. ✅ Post-processing and Validation: The predicted codes may undergo post-processing to ensure they adhere to coding guidelines and standards. Validation might also be performed, either by the algorithm through confidence scoring or by human experts, to ensure the accuracy of the coding. 🎓 Continuous Learning: Many systems are designed to continuously learn and improve over time. As the model is exposed to more data, and as feedback is received from human validators, the model can be updated and retrained to improve its accuracy and efficiency.

Heterogeneous datasets are pervasive today, existing in various domains. Objects within these complex datasets are often represented from different perspectives, at different scales, or through multiple modalities, such as images, sensor readings, language sequences, and compact mathematical statements. Such datasets have been analyzed in the past using Multi-View Learning (MVL), Multi-Task Learning (MTL), and Tensor Learning (TL). In recent years, Multi-Modal Learning (MML) has also been employed. MML is a Machine Learning (ML) approach that integrates and processes information from multiple types of data, with different "perspectives" or "modalities" such as text, images, audio, video, or sensor data. The goal of MML is to leverage the complementary strengths of these modalities to improve model performance and enable richer understanding and predictions. Precision medicine and personalized clinical decision support systems (CDSS) tools have long aimed to leverage multimodal patient data to better capture complex, high-dimensional patient states and provider responses. This data ranges from free-form text notes and semi-structured electronic health records (EHR) to high-frequency physiological signals. While the advent of transformer architectures has enabled deeper insights from merging modalities, it has also required meticulous feature engineering and alignment. In patient monitoring, effectively analyzing diverse physiological signals within CDSS is highly challenging. #MedicalInformatics To address the challenges of analyzing multimodal patient data, the authors of [1] introduce MedTsLLM, a general multimodal large language model (LLM) framework that effectively integrates time series data and rich contextual information in the form of text. This framework performs three clinically relevant tasks (in time-series) which enable deeper analysis of physiological signals and can provide actionable insights for clinicians: • semantic segmentation • boundary detection • anomaly detection At a high level, boundary detection splits signals into periods like breaths or beats. Semantic segmentation further splits time series into distinct, meaningful segments. Anomaly detection identifies periods within the signals that deviate from normal. MedTsLLM utilizes a reprogramming layer to align embeddings of time series patches with a pretrained LLM's embedding space, making effective use of raw time series in conjunction with textual context. They additionally tailored the text prompt to include patient-specific information. Their experiments showed that MedTsLLM outperforms state-of-the-art baselines, including deep learning models, other LLMs, and clinical methods, across multiple medical domains, specifically electrocardiograms (ECG) and respiratory waveforms. Links to their preprint [1] and #Python GitHub repository [2] are shared in the comments.

Most nonprofits say they want better storytelling. What they actually need is better documentation. Here is the difference. Storytelling is the output. Documentation is the system. If your field teams are not capturing consent properly, if programme data is scattered across WhatsApp, if photos are stored on personal phones, if no one logs small operational milestones, no creative agency can fix that at the end of the year. Strong organisations treat documentation as infrastructure. - Clear photo and video protocols - Shared drives with naming conventions - Monthly impact logs from programme teams - Consent formats in local languages - A communication person looped in from day one This is not glamorous work. But it reduces panic during donor visits. It strengthens annual reports. It improves proposals. It protects dignity. Good communication is rarely about talent. It is about systems. If your organisation feels like it is always scrambling before reporting season, the issue is not creativity. It is workflow. And workflow is fixable. . . . . #Documentation #SocialSector #Communications #CreativeAgency #SimitBhagatStudios