OCR App Scanning For Medical Invoices & Reimbursements in Germany

Pia Health is a mobile app for privately insured individuals in Germany. It simplifies managing medical invoices: scan, organize, and submit invoices directly to your health insurer or state aid. Less paperwork, clearer visibility into your healthcare expenses.

All records in one place. Pia calculates your total medical costs over the year. Once your invoices reach the threshold relevant to your contract – exceeding the Beitragsbemessungsgrenze or hitting your annual deductible, the app notifies you when to submit your claims.

Pia digitizes invoice processing for health insurance reimbursements. Users scan physical invoices with their smartphone camera, enhanced by a contrast filter for better legibility. Multi-page scanning is supported.

The captured images are analyzed by Optical Character Recognition (OCR) algorithm, which extracts only the metadata, that is, key fields like recipient name, invoice amount, invoice date, due date, and banking details (IBAN).

This metadata is then processed through a locally-hosted Large Language Model (LLM), which further parses and organizes the data. All this processing is performed on servers that are hosted in Germany. No US-based Services were used. OCR and LLM analysis are essentially powered by open source solutions.

The structured output is displayed in a clear, spreadsheet-like interface. Pia then compiles the data into a clean PDF and attaches it to a pre-filled email with the user's contact information and insurance number – ready for direct submission to their health insurer.

No manual form-filling, no physical mailing, no post office visits. All communication happens digitally via email, compatible with German private health insurers nationwide.

Pia strictly adheres to GDPR regulations. OCR and LLM operations are performed exclusively on servers located within Germany, with no data transmission to third-party services or servers outside the EU.

Both OCR and LLM components use open-source software, running entirely in-house. This avoids high licensing fees and ensures no external parties access sensitive medical or financial data. Full data sovereignty.

In short: Pia is a privacy-first, fully automated health invoice submission system for the German market – built on smartphone OCR, LLMs, and open-source tech, strictly within regulatory bounds.

Once approved, the team integrated both OCR and LLM algorithms into the web app. These integrations added automatic document parsing and metadata extraction from scanned invoices.

The mobile app, at the time developed, used the native smartphone camera without any specialized document scanner. At this stage was the development team’s first iterations in re-evaluating the app’s features and internal implementation despite the app’s functioning. But, at this stage, after internal testing, our B2B client found out a key limitation: image quality from standard smartphone photos, especially without optimization, was insufficient for accurate OCR results.

OCR is highly sensitive to image clarity. Low lighting, poor focus, small fonts, and low-resolution cameras all degrade recognition accuracy.

From a product and technical standpoint, we, as a development team, had to identify the most effective OCR and LLM integrations that offered cutting-edge performance, were fully compliant with GDPR regulations and cost-efficient. This required picking an OCR capable of handling document scan variability while being either open-source or licensed for local, on-premise deployment within EU-based infrastructure. At the same time, choosing an appropriate LLM involves balancing performance accuracy, context length, and processing speed with the constraints of in-house hardware and limited budgets. 

Commercial models like ChatGPT are hosted externally and not viable under GDPR for this use case. The core challenge: delivering the best results in the shortest processing time, using available resources, without compromising compliance.

Open-source LLMs are evolving rapidly. The backend infrastructure must remain adaptive – regularly benchmarking new models for faster or more accurate results with the same or lower computational overhead.

We continuously ensure the technical stack stays competitive and efficient. This includes testing OCR and LLM results against real use cases, assessing cost-to-performance ratios, and advising our client on when to upgrade components.

This improvement allowed for a much better OCR accuracy, even under suboptimal conditions. Better OCR output led to improved LLM parsing performance, with data extraction accuracy reaching > 90% and text processing times of less than 20 seconds, depending on document volume.

Documents are now automatically recognized, contrast-enhanced, and lighting-corrected. Manual entry is largely eliminated. Key metadata fields like IBANs are mathematically validated and marked with a confirmation check. Users then select the intended recipient.

Nevertheless, while the OCR and LLM pipeline achieves high accuracy, users are still prompted to review and verify extracted data. A confirmation screen then follows that allows for manual corrections or additional input to ensure complete accuracy before submission.