Thesis with Aurora Innovation

The Aurora teleQ service includes a statistics-based scheduling support system designed to optimize how healthcare professionals are available for patient interactions via digital channels. By enhancing this scheduling support with an AI-based model, we believe that scheduling can be further improved for both nurses and patients. The goal is to ensure that as few healthcare staff as possible are scheduled, while still maintaining sufficient availability so that patients receive responses within a reasonable timeframe.

As Generative AI (GenAI) technologies mature, new products are emerging in specialized domains that rely on precise vocabulary and demand high-quality outputs. When developing systems that integrate multiple GenAI components in non-trivial ways, evaluating overall system quality becomes complex. Traditional direct comparison methods are often insufficient since the assessment must focus on semantic equivalence rather than word for word similarity. Key challenges include determining whether changes in model choice,  composition of models, model settings, or prompt design genuinely improve quality or inadvertently degrade it.

This proposes a thesis about the process for generation of a sufficiently high volume of synthetic voice data over a noisy telephony channel, simulating conversations between patients and healthcare. Using this data, the work aims to devise a statistically rigorous framework for evaluating the quality of a complex GenAI system, ensuring both robustness and reproducibility of results.

Primary healthcare professionals are increasingly burdened with administrative tasks submitted by patients, often through voicemail systems. These requests, such as appointment cancellations, rescheduling, or prescription renewals are processed within the same channels used for medical triage and medical advice. The high volume and unstructured nature of such tasks create inefficiencies, as professionals must manually interpret, cross-reference, and act on information in different systems that could be semi-automated. This workload diverts attention from direct patient care for an increasingly burdened profession.

This thesis proposes the design of an agentic AI system that processes patient voicemails, analyzes sentiment and intent, and enriches requests with structured data drawn from electronic health systems via standardized interfaces such as FHIR and HL7. The system would generate actionable, information-rich task summaries that medical professionals can quickly review and either approve or deny. Upon approval, the corresponding actions (e.g., updating appointments or renewing prescriptions) would be executed within the electronic health record system. The research can explore accuracy, efficiency gains, and the impact on the health care professional’s work satisfaction.