This three-year PhD project aims to design, develop, and rigorously evaluate one or more proof-of-concept (PoC) systems for a conversational health agent, provisionally titled “Doctor Assistant”, intended to support physicians in routine clinical practice through AI-augmented documentation and decision support.
The project focuses on the conception and implementation of interactive software systems that integrate advanced artificial intelligence models and components to assist clinical activity, with particular emphasis on medical scribing. The proposed systems will automatically transcribe clinical encounters and subsequently extract, structure, and normalize clinically relevant information, including anamnestic data, physical examination findings, descriptions of signs and symptoms, diagnostic hypotheses, alternative intervention options with associated benefits and risks, therapeutic and dietary regimens, and short- and medium-term procedural plans.
Extracted information will be transformed into structured and semi-structured clinical artifacts, such as encounter summaries, formal reports, progress notes, care diaries, operational to-do lists, and follow-up question prompts. Beyond documentation support, the system will implement consistency checks and semantic validation procedures against authoritative scientific knowledge bases, clinical guidelines, and other relevant documentary sources. These checks will be performed at the level of individual cases and patients, enabling the identification of omissions, inconsistencies, or potential deviations from evidence-based recommendations.
From a technical perspective, the project will explore alternative architectural solutions, including modular and agentic pipelines, retrieval-augmented generation grounded in curated medical corpora, and hybrid neuro-symbolic approaches to enhance controllability and traceability. Particular attention will be devoted to issues of accuracy, robustness to acoustic and linguistic variability, calibration of extracted information, transparency of transformations from raw dialogue to structured data, and integration within existing electronic health record (EHR) systems and clinical workflows.
The PoC systems will be evaluated along multiple dimensions relevant to human-AI collaboration in healthcare, including documentation quality, time efficiency, cognitive load reduction, error prevention, clinical coherence, perceived reliability, safety, usability, and deployability in real-world settings. Controlled studies and field evaluations will assess technical performance but also the impact on clinicians’ work practices, decision-making processes, and overall quality of care. The project aims to contribute a methodologically grounded and clinically responsible framework for AI- supported medical documentation and hybrid clinical intelligence.