Trained neural networks automatically deciphered practitioners’ handwriting on millions of medical claim forms
|One of Belgium’s largest health insurance funds and one of Moonoia’s biggest clients, Partena Group needed to scan, recognize, extract and validate data from millions of handwritten medical claim forms which are very difficult to read even by humans. With doctor’s handwriting being a real challenge, patient data, pathology data and payment data make this claim form a particularly sensitive document regarding privacy and security.||Moonoia devised and developed a docBrain-powered document recognition and data extraction solution that was able to decipher medical practitioners’ handwriting based on both traditional pattern recognition methods and neural network technology.|
The project was awarded the “Project of the Year 2017” prize by the UK-based Document Manager magazine.
In the end, the entire document process was 80% automated and optimized across departments, data was transferred automatically into the client’s business systems and only a minimum amount of forms was routed to a quality assurance and management by exception routine for human examination, recognition and data entry. The client now processes around 7 million documents per year.
|This US-based provider of payment processing and information management services to the vehicle fleet industry had a problem with limited manpower that was manually processing around 800.000 invoices every year. Several thousands end clients send their invoices together with all sorts of documents attached. The latter can be single or multi-page and can include virtually any kind of unstructured content: rate confirmation, bill of lading, proof of delivery, receipts, packaging slips, assignment schedules etc. Moreover, these documents are generated by different capture tools, from smartphones to professional scanners, resulting in a large variation of image quality and size.|
A team of forty people is dedicated to the time-consuming task of checking the presence of all required documents as well as validating the data for the transportation of various types of goods or for client reimbursements. Errors are quite frequent, leading to quality issues and poor service. Different document reading solutions were explored and tested, but none was able to overcome the complexity of processing bulk documents. Given that a grand total of 3.6 million images have to be processed every year, adding manpower to accommodate growing volumes was not the right solution. The requirements were to automatically process, validate and approve these invoice packages.
|Together with its partner IN-RGY, Moonoia developed a neural network-based solution that runs on a private Google cloud and does the following:|
IN-RGY added business rules to complete the validation. Packages seemingly invalid were routed to the docBrain quality control and exception management tool for manual validation / rejection by exception.
The main benefit of using Moonoia / docBrain technology resulted in a new streamlined process, more accurate data and on-time invoice payments, as well as the reduction of the processing cost of minimum 40%.
|This bank and insurance multinational customer wanted a solution a KYC (Know Your Customer) process which required the validation of ID images and data captured from various sources – smartphones, personal scanning devices, email attachments, third party business process outsourcing (BPO), etc. – and received in various formats like .pdf, .tiff, .jpg and others. This periodic verification process of customer data authenticity requires compliance with banking sector rules and anti-money laundering, anti-bribery and anti-corruption due diligence regulations. The customer’s objective was to streamline its onboarding and KYC processes with efficient document reading and data verification, easily integrated into the value chain.||The docBrain features / capabilities read multiple document formats of varying quality from multiple angles. The customer business process was changed and automated to include a docBrain document qualification intervention. The docBrain neural network reads, extracts and qualifies the data (find multiple ID, verify the recto/verso, avoid old format, etc.) to be forwarded back to the customer for validation in their business systems. The benefits to the customer are a faster and improved KYC process as well as time, cost and manpower savings since no data entry or re-entry is necessary.|
The customer intends to further use the docBrain platform to conceive new tailor-made solutions swiftly for other business processes to be improved.
|This oil industry customer wanted to unlock hidden data from handwritten historical well drilling records. Hidden data includes depth of drilling and the quality of the underground layers. These archived documents were scanned into images. The images contain handwritten notes as well as depth and well drawings done by hand. The company wants to research this data to assess the feasibility and the success probability of future drilling in certain areas.||Moonoia proposed the docBrain AI-powered recognition technology to address these very specific needs. A dataset of around 6000 documents was used for training the neural networks. Handwritten text was examined for predefined keywords to be transformed in WITSML data.||Moonoia showed that the expected results can be produced without the need for data science or business intelligence. Moreover, improved research results from complex interrogations were achieved without with significantly less manual data extraction.|
In time, the results of this project / use case will lead to a docBrain evolving solution to facilitate easy research on other kinds of documents / data in completely different fields.
|An Emirati government entity aims to extract data from historical legislative documents which include handwritten Arabic text. The documents have often insufficient quality to be automatically read with traditional recognition engines.||Using docBrain custom neural networks, document image quality has improved, resulting in better readability. docBrain is currently training a new neural network model for automatic recognition of Arabic handwriting. The objective is to extract accurate quality data with fewer errors and decrease the number of manual interventions and validations.|
This project illustrates that docBrain can teach itself the Arabic characters (and for that matter any type of non-Latin alphabet) if there are enough documents and data for training the neural networks.
Sectors such as finance, legal, manufacturing, transport, energy and healthcare are seeing their largest companies implementing robotic process automation to optimize business processes and speed up digital transformation. Most of the current-day automation is directed towards redundant, repetitive tasks such as copy-pasting, launching applications, uploading, downloading or backing up emails. This is why more and more RPA solutions include AI technologies, making day-to-day operations not only more efficient but more value-generating.
At the other end, more and more customers are embracing mobility, making smartphones the new document scanners. This raises new issues for those extracting and analysing the data. On one hand, input images are low quality and unstructured, making validation more difficult. On the other hand, validation needs to be instantaneous and reliable so that the transaction is just as quick as snapping a picture.
KAPTAIN is a solution and an ecosystem. It combines the docbrain’s uniquely powerful – and cognitive – data capture with mobile application development and enterprise-level robotic process automation. The result is an end-to-end, mobile, intelligent and automated document processing solution.
Users are prompted via the Shoot and Prove mobile app to send a certain document – ID, proof of purchase, utility meter reading etc. Sending means simply snapping a photo with their smartphone camera. The app automatically adds a time stamp and location to the photo, turning this action into a transaction and generating an electronic original of the document with legal probatory force.
The image is sent to a Moonoia server and docbrain instantly processes the image, automatically extracting the data from the image using specially trained AI models. Processing raw images from customers is a particularly difficult task for machines because prior to extracting the data they need to determine where exactly in the picture is the requested document located. Most of the times, end-customers send images that are not optimized for capture – sizes and formats vary a lot, documents are skewed, lighting is bad and the entire picture is blurry or contains too many other elements besides the document. docbrain applies a combination of image enhancement, intelligent cropping, advanced OCR and many other document analysis engines to be able to automatically extract the requested data from a document. If the content is too difficult to read automatically, Moonoia sends the exceptional case to a dedicated manual validation team.
These workflows can generate massive volumes of incoming data every day, potentially overcharging any company’s backend systems. This is where robotic process automation comes into play, orchestrating new processes for both back and front office, improving the efficiency of large business.