Client and Business Goals

Parata Systems, founded in 2001, is one of the leading providers of automated adherence packaging providers. They support a variety of formats, including multi-dose blister cards, pouches, and vials, to create personalized medication regimens for patients, with the objective of simplifying medication management among patients and improving overall medication adherence. Stringent quality checks in the form of accurate pill detection to ensure the right medications are placed in the right pouches is extremely critical for the success of adherence packaging.

Parata uses its PERL Pouch Inspector in the quality control stage to identify the drug content in each pouch and has the ambition to make PERL the fastest pouch inspection technology in the market today. With this objective in mind, Parata approached Finarb to create an AI-enabled solution that would be at the heart of PERL pouch inspector to improve, tighten, and expedite the pill detection process in adherence packaging.

What is medication adherence packaging?

Approximately 55-60% of global patients struggle with medication non-adherence, impacting health outcomes and costing healthcare providers and drug manufacturers around $600 billion yearly. Medication adherence packaging, which organizes drugs by dose, date, and time, can enhance adherence, thus improving patient health and reducing losses.

Finarb's Computer-vision enabled Quality Control for Medication Adherence Packaging

The AI-enabled Pill detection and verification system designed for Parata can be broadly classified into:

1. The Helios Pill Detection Platform - Computer Vision Model for Verifying Pills inside a Packaged Pouch
2. Opticam Pill Sanity Check - Computer Vision Model for Identification and Classification of Broken or Damaged Pills

Helios Pill Detection Platform: Deep Learning model to verify the contents of the Drug Pouch

Finarb designed the Helios pill detection platform, based on the precise object detection and image classification algorithm to identify, categorize, and validate the contents of a pouch to ensure customers receive the correct prescribed drugs.

The workflow involved using pill images and configuration data from the PERL adherence packaging machine and running deep learning model inference on these images. The model enabled accurate decision-making on whether each pouch is packed with the precise medications. This was done through the Faster RCNN model for Object Detection and the Resnet architecture for Image Classification.

Opticam Pill Sanity Check - Deep Learning Model to Classify Damaged and Accurate Pills

The Damaged Pill Detection or Pipette platform utilized precise object detection and image classification algorithms to detect and categorize the pill as “Broken/Damaged” or “Accurate” before being packaged, to ensure customers do not receive faulty/broken/damaged drugs.

The workflow involved a high-resolution camera placed around the pipette to capture images of the pill, sending the images for preprocessing with further detection using the Faster RCNN model. The cropped image of the detected pill was then sent to classify the pill as “Broken” or “Accurate” using the Resnet architecture. Based on the output of the model, the damaged pills were segregated and discarded, and accurate pills were sent for packaging.

Accuracy of the AI Model

The model, after fine-tuning and pilot testing delivered an average accuracy of 98.6% in the precise detection of pills according to the shape, size, and colour, and classified the pill based on NDC.

The model also detected damaged pills  and extraneous materials with a recall score of 0.929 and an F1 score of 0.963

Scalability of the model

The model is now being trained to identify larger sets of NDC’s with integrated image augmentation technology to increase the accuracy and precision of object identification and categorization to close to 99%. Through distributed computing, we have also expedited the AI training and processing time to scan these large datasets. To optimize the usage of cloud resources and minimize training time without compromising on accuracy of the model, we created a novel algorithm to curate the training dataset such that we include only those training samples that lower the validation loss. This algorithm is based on Bhattacharya distance between older data points and new data points and concordance of the data points in the original model.

Patent filed

Parata and Finarb have filed a patent for this pioneering methodology at United States Patent and Trademark Office in June 2022. The patent has been filed for Finarb’s innovative and pioneering process and workflow through which pill detection accuracy was much higher than industry benchmark (Patent publication #20220261982 and #20220254172)

Business Outcome and ROI

Automating this check through AI models has resulted in 45% higher efficiency, lower cost per package allowing the client to scale their adherence packaging as demand grows. A higher precision in detection of defective pills also enabled improved customer satisfaction.