Innovative artificial intelligence for practice management in medical healthcare

Le résumé fourni par la source

One of the challenges to increasing access to quality healthcare and improving public health is excessive bureaucracy. For instance, the 2021 ‘Assessment of the State of the Healthcare System in Poland’ report by the Supreme Audit Office (NIK) identified several bureaucratic issues, such as complex procedures, system inconsistencies, excessive forms, and unclear criteria and requirements. These lead to wasted time, staff overload, delays in diagnoses and treatments, and lower-quality healthcare. Medical personnel spend significant time on documentation at the expense of patient care. Artificial intelligence (AI) and machine learning (ML) have shown significant promise in the medical healthcare sector.1 Artificial intelligence models can be used to effectively analyse medical data and support physicians in decision-support processes.2 Examples of clinical decision support systems based on ML include the explainable classification of lung diseases in chest X-rays.3,4 These systems leverage deep learning techniques, such as convolutional neural networks, to detect conditions like pneumonia, tuberculosis, lung cancer, and COVID-19. Explainability is achieved through heatmaps, saliency maps, or attention mechanisms, increasing efficacy and decreasing time needed by radiologists to interpret the images. One exciting direction is to develop innovative AI-powered systems for optimizing medical practice management. A graphical representation of a proposed workflow is shown in Figure 1. By enhancing efficiency, these systems can potentially improve access to high-quality healthcare services and support the well-being of both patients and medical professionals. Several factors drive the relevance and potential for the adoption of innovative AI systems: Rising healthcare demand: An aging and growing population increases the need for comprehensive medical documentation. Legal requirements: The law mandates detailed and accurate medical documentation. Healthcare efficiency and quality: Detailed medical notes are essential for effective team communication that is critical for proper disease management. Convenience and mobility: The expansion of telemedicine and mobile clinics necessitates adaptable documentation solutions. Time savings: Artificial intelligence and automatic speech recognition (ASR) streamline communication, freeing doctors to focus on patient care. Precision: —Artificial intelligence-driven notes can reduce human error, enhancing documentation precision. Existing ASR and NLP solutions are more effective for English and German than other languages, especially in medical settings. Therefore, one key challenge is developing an effective end-to-end process for language-specific medical terminology. The ASR model can use existing language-specific ASR systems (e.g. Kaldi and Sphinx), but these systems will need customization and correction for medical terminology. This requires rule-based systems, fine-tuning with collected data, and machine learning-based language models. Language-specific models (e.g. POLBERT for the Polish language, T5) can be partially leveraged but require thorough analysis to address medical errors. It is necessary to build language-specific models to summarize medical interviews and classify medical conditions. Large language models have seen a significant adoption and quick improvement in many domains. Among these, LLaMA has been shown to be beneficial in the medical field.5 Utilizing LLaMA as the foundational model and optimizing it through low-rank adaptation (LoRA) on 236192 MIMIC-IV discharge summaries, has shown to accurately predict diagnosis-related groups for hospitalized patients. Moreover, DeepSeek-R1,6 a model trained via large-scale reinforcement learning without supervised fine-tuning, demonstrated remarkable reasoning capabilities. One possibility is to adopt both LLaMA and DeepSeek-R1 as pre-trained models, and fine-tune them on medical data to better fit domain characteristics. To account for the limitations of large language models as they encounter complex and domain-specific language in the medical domain, one opportunity is to leverage open-source tools such as LangChain,7 widely recognized in the AI community for its ability to seamlessly interact with various data sources and for its modular abstractions and customizable, use-case-specific pipelines. Alternatively, adopting the retrieval-augmented generation (RAG) technique,8,9 which combines pre-trained parametric and non-parametric memory for language generation, is a viable option. While advancements in medical ASR and language models are ongoing, many healthcare systems lack comprehensive solutions tailored to specific languages and regional medical practices. For example, existing ASR research efforts have yet to produce a system capable of accurately converting language-specific medical interviews into structured notes. Future AI development should prioritize language adaptation to ensure inclusivity across diverse healthcare environments. Artificial intelligence-based transcription should complement rather than replace human expertise. Physicians play a crucial role in validating and refining AI-generated summaries. A physician-in-the-loop approach ensures that AI suggestions remain clinically relevant, enabling doctors to review, and correct generated notes rather than drafting documentation from scratch. Physicians will continuously refine the AI model, ensuring it adapts to evolving medical terminology and emerging procedures and techniques. Artificial intelligence-driven transcription solutions should undergo rigorous optimization, testing, and integration into healthcare workflows to ensure practical usability. The key steps in this process include: Evaluating ASR models in clinical environments and curating high-quality training datasets. Optimizing AI models for efficiency, accuracy, and interpretability. Deploying and testing AI solutions that align with existing real-world healthcare infrastructure while maintaining transparency and compliance. The ultimate goal would be to offer a significantly improved AI system for medical practice management including: Automatic creation of medical notes using AI. Automatic medical notes for doctors and patients. Automatic coding suggestions (ICF, ICD-9, and ICD-10). The output of the first step can be provided as input for the second step. This technique involving RAG addresses possible limitations of LLM as they encounter domain-specific medical terms. These technical features would lead to improved healthcare-related outcomes: Improved documentation efficiency, reducing administrative burden for healthcare professionals. Enhanced diagnostic accuracy through structured and AI-assisted clinical documentation. Better patient and medical team communication via automated, personalized summaries. Optimized billing and compliance, ensuring accurate coding and reimbursement. Increased accessibility to high-quality care by streamlining medical workflows. Proposed workflow of innovative artificial intelligence methodology Possible directions include the adoption of patient data in AI models for real-time monitoring and preventive diagnosis. Considering the possibility of mistakes and hallucinations that AI models may be subject to, we envision physician-mediated interactions as the most viable way to exploit the technological value of AI while providing medically sound advice to patients. All authors declare no disclosure of interest for this contribution.

Ce résumé expose les affirmations des auteurs. BNTIC ne l’interprète pas comme une validation indépendante des résultats.