PillPoint

Medication Classification vs. Amount of Images used
Data between training and testing sets for Deep Learning
Testing Data Values
Training Data Values
Deep Learning Learning Progress
Pill Sorter
Pill Tray
Housing Unit

Our project, PillPoint, is an augmented drug identifier and sorter for automatic pill dispensers. Automatic pill dispensers are not a novel concept and already serve as a great resource for those who take many medications daily, but we feel there are some issues with them that could be tackled. The largest is that most of these dispensers require the pills to be very carefully loaded in, and for the groups for whom these dispensers are targeted, namely the elderly and people with conditions affecting memory and fine motor skills, loading in the medications may prove a difficult barrier. In addition, we feel that the design of most pill dispensers is too reminiscent of laboratory equipment and could take on a more subtle, kitchen-appliance-like aesthetic. In our device, every pill inserted will be analyzed to determine its brand. The patient can pour in an entire month’s worth of pills all at once, where they will be automatically identified and sorted based on pill dimension and markings using image recognition. The patient would then input what times they want to take their medication(s) so that the PillPoint can dispense and alert them at those times.

In training our model, we used the Deep Learning Toolbox and Image Processing Toolbox from the MathWorks Suite. We trained our model using a dataset of 1,322 images which consisted of twelve medications in their numerous pill/capsule varieties, and had the model use deep learning to analyze the characteristics of pills from the dataset with the goal to correctly identify each drug. In addition, we provided a design for the PillPoint that uses a funnel to intake the pills and delivers each dose via a slide and tray delivery system. One thing that definitely needs work is our model— in the limited time we had, we were able to train a model with 45% accuracy. For PillPoint to be a viable product, accuracy would need to be incredibly high to avoid any liability issues. We also struggled to assemble the program to schedule dose reminders. Despite the lower-than-ideal accuracy rate, the experience taught us a lot about using Deep Learning Software, and we feel that we made some important steps forward in tackling this issue.