Digital patient referral form that must be filled out.
General outline of the machine learning process.
Overview of the new referral process. Red arrows represent the user and system feedback.
What it does
AiRefer utilizes the referral history of clinics to predict the appropriate specialists for a patient to see based on urgency, proximity, and the subspecialty of the clinician allowing the referring doctor to book an appointment at the time of diagnosis.
One of the most prevalent issues in healthcare is providing care to patients in a timely manner. Many have suffered from long wait times and not being seen within the appropriate urgency protocol. Currently, physicians issue a referral by faxing profiles of a patient. From there, the specialist office must schedule an appointment according to the severity of the symptoms. It can take days for the specialist to schedule an appointment, and oftentimes, patients aren’t seen within their required urgency date. Our goal was to create a system which can generate an optimal referral in seconds, reducing inefficiencies within the current system.
How it works
AiRefer will have an electronic referral sheet to be filled by the referring doctor. This sheet will hold information such as patient symptoms, symptom duration, and how far they are willing to travel. Using machine learning, an artificial neural network will be trained for regression to determine the urgency of the patient’s condition based on symptoms and duration. The urgency will be used with proximity and subspecialist availability in a second neural network for classification. This second neural network will output a list of optimal specialists, and it will predict the appropriate appointment dates by anticipating the future referrals for each clinic. Specialists can then confirm the referral and the patient will be notified. Currently, the neural networks are being trained by patient data acquired from a local hospital. Since a deep learning approach is being used, the accuracy of the algorithm will increase the more it is used.
As with any good design, AiRefer started at the brainstorming table. Through many iterations and modifications, the idea of revolutionizing the medical referral system was conceived. Prototyping started with a low-fidelity diagram that laid out the structure of the idea. From there, we contacted local medical, academic, and industry professionals seeking guidance and feedback. Once most of the conceptual holes were filled, sample referral data was obtained from a local clinic, and processing commenced. AiRefer is being built in stages to ensure a small scope is maintained and any issues are swiftly addressed. The first stage consists of optimizing the algorithm to predict the urgency of a referral so that it can later be scheduled. The back-end framework for referral urgency has been coded and the algorithm is currently being optimized to attain accurate predictions.
How it is different
AiRefer is unique as it steps away from the current referral model. Rather than having referrals be faxed or sent to the specialist’s office for administrative personnel to book an appointment, AiRefer uses objective machine learning algorithms to make accurate predictions. Although different methods have been suggested and employed to reduce wait times for specialists, many depend on transferring the scheduling to a different human to triage. AiRefer is different as it is able to constantly and objectively learn from the most recent data, so that accuracy is maintained.
The next stages of AiRefer’s development will consist of obtaining additional data and building a user interface so that testing can begin. We have partnered with a clinic at the University Hospital so that specialists and referring, emergency department clinicians will be able to test the system and provide feedback. From there AiRefer will expand to work with multiple clinics of the same specialty so that a network can be built. Subscribing clinics will have their referrals processed by AiRefer and local family doctors can start using the system.