ML Algorithms to Identify Dental Insurance Fraud with 95% Accuracy

About Our Client

The Client is an insurtech startup focused on innovative software solutions for dental insurance management.

Data Science Skills Needed to Create Insurance ML Algorithms

The Client was developing a cutting-edge software product for identifying dental insurance claim errors and fraudulent submissions. According to dental insurer sentiment, dental care providers often made mistakes when filling insurance claims (e.g., listing a wrong tooth code). There were also cases of intentionally changing the diagnoses and submitting claims with the same or manually altered X-rays multiple times to receive higher reimbursement.

The Client’s software product was meant to use computer vision and machine learning (ML) to process reimbursement claims and attached X-rays, analyze insureds’ dental conditions, and compare them to the diagnoses and procedures listed in the claim. In addition, the software was to match the received X-ray images against the images uploaded to the insurer’s database, spot duplicates and manual image modifications, and report fraud.

The Client needed robust data science and ML engineering skills to build advanced computer vision algorithms for dental image processing, oral health condition analysis, and discrepancy detection. The Client was looking for a trustworthy vendor that could promptly provide the required competencies to accelerate the solution’s time-to-market. Owing to ScienceSoft’s decades-long experience in ML implementation and engineering custom medical image analysis solutions, the Client turned to us for a seasoned ML expert.

Development of ML Algorithms for Dental Insurance Fraud Detection

To meet the Client’s needs, ScienceSoft allocated a full-time Senior Data Scientist / Senior Python Developer with team lead experience and deep knowledge of both health insurance and healthcare specifics. Within a week after the initial request, ScienceSoft’s expert passed the knowledge transfer and was ready to join the project. The expert analyzed the available project documentation and collaboration flows and started working on ML algorithms along with the Client’s data scientists.

The joint team delivered the ML-fueled computer vision algorithms for:

• Identifying tooth type (incisor, canine, premolar, molar) and location (coded using a standard tooth numbering system).
• Detecting dental problem location within a tooth.
• Recognizing the oral health condition type (cavity, decay beneath the existing filling, abscess, root rot, cyst, etc.).
• Validating the authenticity of dental X-ray scans via comparing the submitted medical images to the images from the insurer’s database.
• Identifying duplicated medical images and image alterations (e.g., image contrast tuning, adding new image elements using editing programs, image alignment changes) to prevent fraudulent practices exploited by dental care providers.

ScienceSoft’s expert applied his 10+ years of experience in engineering convolutional neural networks (CNNs), an advanced class of computer vision models employed in the Client’s project, to deliver reliable algorithms and effectively solve arising technical challenges. In particular, the expert created background algorithms for unifying the different radiology information systems’ (RISs) formats of dental X-ray files to establish standardized image processing flows. Collaborating closely with the project team, the expert also tackled the issues with X-ray quality, sizing, alignment, and irrelevant data mentioned on the scans.

To address regulatory requirements for the transparency and explainability of intelligent claim processing algorithms, ScienceSoft’s expert suggested using LIME and SHAP techniques to interpret step-by-step ML model logic. For image analysis algorithms, the expert applied class activation mapping (CAM) and feature maps. CAM technique enhanced output interpretability by showing which parts of the input dental X-rays contribute most to the model’s classification logic. Feature maps allow tracing the features (edges, textures, etc.) learned by the image analysis algorithms at different X-ray layers to understand how the model transforms raw images into meaningful representations for fraud detection.

The Client planned to promote its insurtech product as Software as a Medical Device (SaMD). An ISO 13485-certified vendor, ScienceSoft followed the structured quality management practices outlined in this standard to deliver a high-quality solution.

Achieving 95% Accuracy of Fraud Detection ML Algorithms

After the 6-month-long engagement of ScienceSoft’s Senior Data Scientist, the Client got a market-ready MVP for its dental insurance fraud detection solution. On average, the algorithms demonstrated a 95% accuracy rate in detecting dental conditions and inconsistencies in submitted claims, which illustrates the technical robustness of the solution and positions it as a reliable tool for efficient dental insurance claim processing. After the development was finished, the Client submitted the product to the FDA for clearance.

The Client’s in-house development team appreciated the profound ML expertise, proactive approach, and impeccable communication skills of ScienceSoft’s expert.

Technologies

Python, Pytorch.