For this specific project, we had to gain a thorough understanding of the many challenges that this build faced:

Document volume – MRC focus on clinical negligence and personal injury cases. Due to the complexity and variety in these cases, some case documents could span thousands of pages in total, whilst others may consist of 50 or less.

Classification variety – To detect handwriting from doctor’s notes as well as specialist notes for personal injury consultations. It would have to do this as well as quickly adjusting to typed copy. The quality of each document also becomes a factor when classifying blurry and low-quality scans.

Value differentiation – To build accurate models, there is a need to understand and differentiate between numbers and characters that shared similar visuals. For example, ‘1’ and ‘7’, or ‘6’ and ‘G’. Failure to differentiate numerical and character differences would result in unstructured page formatting and classification for indexing purposes.

Data modelling – A complex variety of label types to identify such as hospitals, patient details, medical procedures, and processes.

Due to the complexity of the build and the requirement for validation to the business and its stakeholders, a Proof-of-Concept (PoC) focused on the above-mentioned challenges was to be delivered.

To combat differentiation, the PoC we built had to assign a category/subcategory for each document as well as understand the classification variety. This would help replicate the current system and requirement used by MRC through their manual process.

To begin streamlining MRC’s operation and collation processes, we worked on building a PoC that could fully automate the document processing, sorting, and collation system. Using Amazon Textract as the core service, we could intelligently extract texts from handwritten, scanned, or electronic medical records. Then, by applying Amazon Comprehend capabilities, we could extract insights from the documents such as key information or document category.

We used Comprehend Custom Classifier to manage the data model training and predictive capabilities. A TFIDF (term frequency–inverse document frequency) text classification model was built as a precursor to modelling performance and understanding the data at hand.

These ML (Machine Learning) services were built on a serverless infrastructure using AWS Lambda and Amazon DynamoDB. This provided greater scalability, flexibility, and reduced cost during the data training and concept building. As a result, at the end of this AI/ML driven process, MRC can create a neatly ordered, paginated, and indexed document.