Mental health claims are denied at nearly twice the rate of other medical services. Solo practitioners—who represent 47% of psychologists—often lack the billing staff, legal teams, or infrastructure to effectively handle appeals. As a result, patients are left to navigate complex policy language and administrative processes on their own.
We saw an opportunity to build a tool that empowers individuals to advocate for themselves with the support of AI, precedent, and payer-specific policy guidance.
AppealRX helps patients rewrite and improve insurance appeal letters after a denial of mental health services. Users upload their draft, doctor’s note, and any supporting documents. The system analyzes the submission, retrieves similar past appeals and policy rules, and returns a refined draft with specific suggestions.
AppealRX also includes a success prediction model—a classifier trained on historical appeal outcomes—which scores each draft’s likelihood of being approved. Together, these tools give users both actionable edits and a clear understanding of how their case compares to successful ones.
Frontend:
- Built using TypeScript with Next.js
- Styled with Tailwind CSS
- Deployed via Vercel
Backend:
- MongoDB stores user drafts, model outputs, and metadata.
- Pinecone is used as our vector database, indexing past appeals, insurer policies, and mental health guidelines.
- A fine-tuned BERT model trained on 7000+ appeals notes classifies whether a draft appeal is likely to result in a favorable decision
- A custom Retrieval-Augmented Generation (RAG) pipeline retrieves relevant legal and clinical language.
- To manage vector scaling, we route queries by insurance provider and preprocess large PDFs to extract and index only relevant content.
- Data Quality: Medical appeals data was unstructured and inconsistent. Many documents included the appeal decision in the note body, causing label leakage. We resolved this by lemmatizing and filtering out outcome-related terms like "denied", "approved", or "overturned" and leading alnguage.
- Database Scaling: Retrieval slowed as more appeals and documents were added. We built a routing layer to narrow the search scope based on the user’s insurer.
- Policy Document Parsing: We needed to preprocess PDFs to extract key terms (e.g., definitions of medical necessity) to feed into the RAG pipeline.
- Built a fully functioning RAG-based appeal assistant in under 36 hours.
- Fine-tuned a BERT classifier to predict appeal outcomes with strong accuracy.
- Designed a scalable vector retrieval system based on user context (insurer).
- Created a clean and intuitive user interface to lower the barrier to appeal.
- Preprocessing is critical when dealing with real-world clinical data.
- Building a fast and relevant retrieval system requires careful routing and vector management.
- The mental health advocacy space is under-supported—tools like this can offer huge impact with even small improvements in usability and clarity.
- Expand to include other medical denial types (e.g., chronic conditions, surgeries).
- Partner with advocacy nonprofits and solo practices to pilot real-world cases.
- Add tools for clinicians to auto-generate appeals for patients.
- Build dashboards to track reasons for denial and provide best practices for future appeals.
Ultimately, we want to make appeals fairer, more transparent, and easier to win—especially for those most vulnerable.
bertdeepseekmongodbnext.jspineconepythonrailwaytailwindtypescriptvercel
