mimilabs

Playing with Anthropic's MCP + mimilabs while waiting at ATL airport Trying to make the most of my airport waiting time, I decided to play around with what's been really popular lately - Anthropic's Model Context Protocol (MCP). For those who heard MCP for the first time, it provides a framework for Claude to use various tools, making it way more powerful than standard LLM interactions. In my case, I wanted Claude to access the detailed column and table descriptions of all the tables curated at mimilabs. So I spent about an hour building a quick integration. Here's how things work (see the video): 1. When a user asks a question that's "seemingly" related to any dataset from MimiLabs, Claude "figures it out" and asks: "Can I use this MimiLabs tool?" 2. Then, with just one click, Claude gets equipped with all the open data context it needs It's pretty cool what you can build in just an hour! I've attached a video showing how it works. It's still under development, but I think I can make this available to the mimilabs members soon! Anyone else experimenting with MCP? What cool integrations have you tried? #Claude #MCP #MimiLabs #AITools #QuickBuild #AirportHacking https://lnkd.in/d4rPGPUs

Regional Utilization Differences in NY Healthcare is local. Demographics are different across different regions. Hospitals and post-acute facilities are also different. The dynamics of local health plans and the providers is different across the regions. We conceptually understand this, but looking at how exactly these things affect in utilization, costs, and qualities can be quite challenging. Today, I was looking at the New York market. Using the CMS Physician Supplier Provider Summary (PSPS) data, which contain the claim payment amounts per HCPCS per carrier, I wanted to understand the differences in the utilization volumes. The chart below shows the total Medicare paid amounts between 2020 and 2023 grouped by BETOS Codes. The color represents the carrier location; Queens, Manhattan, etc. As each region has different volume, I normalized based on the M1B (Established Office Visits). What do you see? I see a few interesing discrepancies between regions. - O1E: The rest of the NY state has way higher J-code utilization than Manhattan and Queens. - M4B: The SNF volume is quite substantial in Queens compared to other parts of NY - P6C: Interestingly, Queens and Long Isaland has relatively higher volume of Physical/occupational Therapies Now the question is "why" Why do we see these differences in one state? And how might these regional utilization patterns impact the implementation of value-based care models across New York? Are we designing payment models that account for these legitimate regional differences? Appreciate any insights from the local market experts. BTW, do you want to see how I did this analysis? The script is available at mimilabs!

Vibe-Analysis for Predicting Next Year's ACO Savings I've always told folks that ACO-level savings are relatively easier to predict than doctor-level or practice-level predictions. Why? ACOs have more samples in the group, and individual-level variations tend to cancel out. Plus, ACO-level savings are affected by more stable factors: growth strategies, regional competition, and regional benchmark gaps. These factors don't change as drastically over time as patients' risk profiles and service utilization. Plus, as some of you know, recently, I've been experimenting with vibe-coding and vibe-analysis lately. So I thought, why not test the limits of this approach with ACO predictions? Here's what I did: 1. Used the dataset from mimilabs 2. Provided context and instructions: predict next year's savings using current and previous year's data 3. Asked Claude to set aside the latest data as a hold-out set to evaluate prediction performance 4. Asked Claude to use mean imputation on a yearly basis 5. Asked Claude to use both ElasticNet and XGBoost models (well, I kinda micro-managed... anyway) The AI generated about 700 lines of Python code, which I reviewed manually (with some help from Claude) - and the quality of the code was pretty good. The results? Using XGBoost, you can actually predict next year's savings fairly well at the ACO level. Turns out that good ACOs tend to remain good, and vice versa! The source code is available for mimilabs members. Take a look and try vibe-analysis yourself!

Tracking the Lineage of Medicare Advantage Contracts Coincidentally, several colleagues have recently asked me if there's a way to track ownership changes of Medicare Advantage plans. At first, I didn't have a clear answer, but after some consideration, I realized we could potentially detect these changes by tracking parent organization name changes over time. (One caveat: not all name changes indicate ownership changes, but this approach can significantly narrow the search space). So tonight, I decided to try this with Claude.ai, and as always, I was impressed with the results. Using ten years of Medicare Advantage directory data from mimilabs, we identified all parent organization name changes and when they occurred. I then asked Claude to analyze the top 3 contracts with the most ownership changes and visualize them (see below). The visualizations revealed interesting patterns for contracts previously owned by Bright Health, Trinity Health, and Universal American. Each underwent different M&A activities and complex ownership transitions over the past decade. What's particularly interesting is that many parent organization changes occurred in 2023 and 2024. Will this trend continue into 2025? Only time will tell. By the way, the SQL script and output data are available for the mimilabs users!

Welcoming Prince Baawuah, DrPH, MPH, CSM? to the team to strengthen our healthcare program integrity mission! At falcon health, our mission is to ensure healthcare programs operate as intended, with minimal fraud, waste, abuse, and data errors. Healthcare payments work in amazingly complex ways, and systems have only grown more layered over time. Fighting FWA is a big part of what we do, but understanding the programs' original intent and keeping them aligned with their intended directions is equally important. This can be surprisingly challenging when there are always those looking to game the system. Neil Ahuja and I have known Prince for some time. As many in the industry know, Prince is deeply passionate about educating others on the complex inner workings of Medicare Advantage—one of the largest government healthcare programs—and has extensive knowledge in this space. We share the same mission: making systems work as intended and simplifying complex concepts so people don't get confused. After numerous virtual and in-person interactions, we realized one thing was clear—we needed to work together! So, please join me in welcoming Prince. He will be leading our Medicare Advantage solutions at Falcon (and also joining as a builder at mimilabs). Falcon, about us: https://lnkd.in/eezB_E3C mimilabs, about us: https://lnkd.in/eWt8RDTN

More than ten thousand queries run on publicly available healthcare datasets at mimilabs! I founded mimilabs to liberate the use of publicly available healthcare datasets. Yes, I wanted people to see the value of these gems, and wanted them to use more! There are a lot of statistics we can pull out, but this one made me happy today. This one shows the number of SQL queries that run on the mimilabs workspace per month. Each color in the chart shows different users. And, in February, we surpassed the 10K queries per month threshold! (it was my secret target). Thanks again to all the mimilabs users, and let's keep digging!