Details

Autor(en) / Beteiligte

• Makthal, Sameer

Titel

Investigating Data Driven Approaches to Analyze Patient Handling of Eye-Drop Bottles

Ort / Verlag

ProQuest Dissertations & Theses

Erscheinungsjahr

2023

Link zum Volltext

• ProQuest

Quelle

ProQuest Dissertations & Theses A&I

Beschreibungen/Notizen

• Glaucoma is the leading cause of irreversible blindness and projected to affect over 100 million across the world by 2040. It is widely understood that glaucoma is caused by the death of retinal ganglion cells, but a lack of understanding of its etiology, and the patient burden associated with daily topical medication for the pharmacological reduction of intraocular pressure (IOP) makes it difficult to diagnose and manage post-diagnosis. This dissertation contributes to the management of glaucoma post-diagnosis through the evaluation of useinspired strategies that leverage data for identifying and managing this disease effectively.A rise in IOP is often suggested as the primary cause of glaucoma and is currently the only factor that can be modified for the treatment of this disease. A reduction in IOP can be achieved through the regular application of eye-drops and monitoring the adherence of patients to their treatment plans. Due to glaucoma being highly prevalent in the older population, adherence to regular application of eyedrops remains a major point of concern. Therefore, this body of work focuses on data-driven strategies based on the behavior associated with patient handling of eye-drop bottles for the effective management of glaucoma post-diagnosis. A deep learning approach to human activity recognition is opted to develop a model with the ability to understand patient behavior associated with adherence to glaucoma medication. Accepting the variety in scenarios associated with eyedrop medication, multiple models were developed to better explore the effectiveness of data and patient behavior for the management of this disease.Three unique scenarios were modeled using long short-term memory (LSTM) recurrent neural networks (RNN). Scenario ‘A’ involved the use of a drop delivery aid called GentleDrop™ to help patients in administering a drop. A model was trained to successfully predict the all involved activities based on inertial data recorded from volunteer interaction with eye-drop bottles. Scenario 'B' addressed the limitation of HAR models to predict the same activities recorded at different sensor orientations. Model 'B' was trained on datasets with transformed coordinates added as an additional feature. Scenario 'C' involved developing a human activity recognition model without a drop delivery aid and using a force sensitive resistor to detect handling force. Model ‘A’ displayed good performance in predicting the action of administering a drop. Model ‘B’ performance was on par with model 1 with specificity percentages over 80%. The performance of model 'C' was better in comparison, exhibiting high precision and specificity percentages. The three models showed good overall accuracy of 88%, 88% and 89% respectively.The use of such models could have a positive impact on identifying the barriers associated with passive attitude towards topical medications. In the future, data driven models could also be integrated with the care provider management ecosystem to actively improve patient adherence to glaucoma treatment plans by intervening when medication usage wanes.