Analysing a Discharge Summary using AI for Automated Claims Decisioning

By Parag Sharma, Himanshu Saraf

The discharge summary is a critical insurance claim document that contains important information about a patient’s hospital visit such as the disease the patient suffered from and the procedure performed on him in the hospital. Typically this information is read from a physical document and manually entered into the system by a claims agent and then a medical expert manually reviews this information (in conjunction with other claim documents) to assess the admissibility of the claim. The manual process is not only slow and expensive but also error-prone. It is also not easily scalable as processing more documents requires more people and infrastructure including expensive QA resources.

Flowmagic not only automates the process of extracting the key data from the discharge summary but also analyses this data to make useful predictions.

Insurance companies use TPAs or have medical coders on their payroll to assign ICD codes to diagnosis and procedures. This manual classification is a labour intensive process that consumes significant resources. Flowmagic has inbuilt sophisticated AI models to determine the appropriate ICD-10-CM codes from the diagnosis and ICD-10-PCS codes from the procedures performed or course followed in the hospital.

Flowmagic eventually predicts if the disease is covered by the insurance product held by the customer.

Key Benefits of using Flowmagic for analysing discharge summaries

1) Works for all formats of discharge summaries
Flowmagic extracts structured data from the discharge summary without a user having to tell the model which fields to extract or what the format of the document is. As long as the document contains printed information in pdf or image format flowmagic can read it.
2) Improves over time
The model logs  the mistakes it makes and uses it to retrain itself so that it is continuously improving and doesn’t repeat mistakes.
3) Can be easily trained
The flowmagic platform provides a simple visual interface for the user to train the model using bulk data to realize significant improvements in prediction accuracy and to keep the model current.
4) Makes Pay or Reject recommendations
In addition to analysing the discharge summary and determining if the diagnosis is covered by the policy or not, the solution can be integrated with other flowmagic applications for end to end processing of a claim and making pay or reject recommendations for all but the most complex claims.
5) Decisioning in a matter of minutes
State of the art AI models working on top of a robust architecture mean that the claims that earlier took days to process can now be processed automatically in a matter of minutes.

How It Works

The flowmagic solution for analysing discharge summary consists of four steps:
Step 1: An OCR extracts text from the discharge summary
Step 2: This unstructured text is processed by NLP based AI models to identify all the relevant fields in the document and their values and stores it in a structured format such as a JSON or XML.
Step 3: AI models predict the ICD-10-CM codes based on the diagnosis data and ICD-10-PCS codes from the procedures performed.
Step 4: Finally the diagnosis information is fed to a product-specific AI model which determines whether the disease is covered by the policy held by the customer or not.



Machine Learning Models Used

A discharge summary has multiple sections such as a header section, a footer section having the hospital details, and a section having the patient's details, and the main body that has vital information such as the diagnosis, investigations, procedures performed during the stay in the hospital etc. Flowmagic segregates this data using multiple binary classification SVM models rather than a single multi category classification model. Using multiple models significantly increases the accuracy of feature detection for classification. This process of separating one feature from another helps in developing a general solution that works on a variety of discharge summaries from different hospitals. The flexibility Flowmagic provides to the user to independently train a model, helps in the model to even evaluate a new discharge summary.

Flowmagic can extract information from discharge summaries with an accuracy of over 90% and can make scope coverage predictions with an accuracy of over 85%.

To learn more about how Flowmagic can help you in automating your claims processing or to discuss your broader AI goals, please get in touch with us at