LC Buddy - Part 3
In this project, I idealized the app 'LCBuddy', a smartphone-based self-assessment tool for postpartum patients experiencing breast pain. This tool is designed to improve telehealth experiences by enabling efficient patient-provider communication through momentary assessment and AI mediation. Its main components include a user interface for acquiring patient pain information and two deep-learning algorithms for image quality assessment and condition classification, aiding in remote diagnosis and triage. The tool guides patients based on their condition and uses a severity-based feedback system to help lactation consultants prioritize cases.
In-depth content analysis and expert review for application design
The study consisted of creating a customized self-assessment tool for postpartum patients suffering from breast pain. This tool is intended to help lactation specialists manage their patients by providing a list of patients who are currently facing breast pain from breastfeeding complications, where they access patient requests, prioritizing the most severe cases. On the patient's side, this tool allows the log of momentary symptoms and provides self-intervention options while the patient awaits a professional response.
Steps taken during the study:
(1) Content analysis of existing comprehensive health history (CHH) patient intake forms and an evaluation of specialized protocols for assessing breastfeeding-related pain, gathering specific questions, symptoms, and any information that is key to identifying the most common problems.
(2) Brainstorming a customized CHH form designed specifically for postpartum breast pain cases, focusing on gathering detailed patient information.
(3) Development of 2 accessory AI algorithms to help check image quality for proper evaluation and for pre-diagnosing the patient's condition based on wound image.
(4) Conduct the expert review to validate the customized CHH self-assessment form, identify the system workflow and the main UI components that should be used in the patient application.
(5) Design and development of a smartphone application using React Native and Python for the self-assessment pain tool (under development and in continuous validation with the expert).
LCBuddy System Overview: Our pain assessment application allows the patient to solicit remote guidance from their provider (i.e., lactation consultant (LC)) by informing details about their pain and providing an image of the wounded area. An AI pipeline analyzes the image for a quality check and pre-determines the possible condition of the patient, reporting this data to the healthcare provider and providing the patient temporary guidance to mitigate the issue while the patient waits for the provider’s contact. At the same time, the LC receives patient reports in order of severity so they know who to prioritize first.
AI Algorithms
I developed two deep-learning algorithms to address the challenges informed by the expert in telehealth scenarios, particularly the lack of guidelines and quality checks in patient-provider image sharing for lactation consulting. Both datasets consisted of images collected from diverse platforms such as breastfeeding-related books, articles, online blogs for mothers and physicians, YouTube from educative organizations, and social media platforms (e.g., Instagram, Facebook, Twitter) from healthcare providers who would have educative resources for mothers.
(a) Image Quality Evaluation. This algorithm focuses on image quality evaluation. It was trained and validated using a self-gathered dataset with 1024 images containing breast images and upper body parts in various conditions, ensuring its ability to accurately assess 9 conditions. Our preliminary results achieved an average accuracy of 98%, with precision and recall metrics averaging 82% and 91%, respectively.
(b) Breast Condition Evaluation. This algorithm classifies images into categories related to common breastfeeding pain conditions. It was similarly trained on a self-gathered 1000-image dataset with seven breastfeeding-related conditions, and its preliminary results demonstrated an average accuracy of 96%, with precision and recall metrics averaging 86% and 82%. This algorithm shows promising results for incorporating preliminary diagnoses into the system, guiding LCs in decision-making. The image below shows the working mechanisms of this model, which classifies the images between seven conditions: No apparent problem, engorgement, mastitis, abscess, dermatoses, nipple bleb and nipple damage. The development of the algorithms is currently under review in the journal JMIR.
Outcomes from this study
The preliminary findings from this study were presented as an extended abstract at CHI'24. The feedback received from the community informed the development of my third dissertation article which is currently under development and will be published soon.