Recenti progressi in medicina最新文献

Introduction: Major pancreatic surgery, such as pancreaticoduodenectomy, often involves patients with compromised nutritional status. Parenteral nutrition (PN) may serve as temporary support while awaiting adequate oral intake.

Aim: This report presents a clinical case to discuss the indications, benefits, and implications of PN within an ERAS-based perioperative management protocol.

Clinical case: A 71-year-old woman with recent weight loss and a BMI of 18.28 underwent a Whipple procedure for distal biliary tract cancer. Preoperatively, nutritional supplementation with immunonutrients was initiated. Postoperatively, PN was administered via a peripheral venous line using a three-chamber bag, without central venous catheter placement. Oral feeding was gradually resumed, and PN was discontinued on postoperative day 6. The postoperative course was uneventful, with no clinical or surgical complications. The patient was discharged on postoperative day 8. Histological examination revealed in situ intrabiliary neoplasia without pathological lymph node involvement, and no indication for adjuvant therapy. At the 14-day postoperative dietary follow-up, the patient showed a weight gain of 2 kg and adequate nutritional intake.

Discussion: In the setting of high-complexity pancreatic surgery, early and individualized nutritional management is crucial, as recommended by ERAS and ESPEN guidelines.

Conclusions: In this case, peripheral PN proved to be an effective and safe strategy, avoiding the risks associated with central venous catheter placement. The integrated approach supported a rapid nutritional and clinical recovery.

Using Sleep Medicine guidelines and textbook, we evaluated four large language models (LLMs) (Llama 3.2 3B, Llama 3.3 70B, GPT 4o mini, Gemini 2.0 Flash) on AIMS certification questions, comparing baseline and Retrieval Augmented Generation (RAG) performance. RAG improved accuracy in all models (e.g., Llama 3.2 +9.6 points, Gemini 2.0 +4.0 points), highlighting RAG's role in enhancing LLM reliability in specialized medical domain.

An AI system based on NLP and machine learning has been developed to identify surgical site infections (SSIs) from hospital discharge letters. After advanced pre-processing and imbalance handling, BERT-FT achieved the best performance (F1=0.79), outperforming TF-IDF and W2V. Large language models (LLMs) showed limitations. The system could support semi-automatic SSI surveillance, with prospects for optimisation in translations, prompts, and infrastructure.

Breakthrough digital phenotyping approach reveals three distinct behavioral patterns from smartphone data that could revolutionize personalized mental health care. Using AI clustering on 77 users, we discovered "Night Owls", "Routine-Oriented", and "Always-Connected" behavioral types with 90%+ accuracy. Our explainable ML pipeline identifies key digital biomarkers for targeted interventions, offering clinicians data-driven insights for precision psychiatry.

This study shows that machine learning can accurately distinguish between mitochondrial and nuclear DNA mutations in primary mitochondrial diseases using only non-genetic and non-histological clinical data. While language models underperform in comparison, they show potential as complementary diagnostic tools.

Clinical research is increasingly regulated. Despite growing artificial intelligence (AI) use in healthcare, there is a lack of adequate tools to support researchers in non profit (AI or not) studies. To assist with the classification of clinical software, ClinEthix, a prototype conversational tool, has been developed to help researchers with regulatory qualification. A survey of 20 researchers found it highly useful, clear and user-friendly. Future developments will integrate LLMs and human feedback to improve accuracy.

Large language models (LLMs) show promise in explicit reasoning for complex medical fields like psychiatry. This study assessed the clinical validity of Gemini's chain-of-thought (CoT) reasoning in 10 complex psychiatric cases, evaluated by specialists using six metrics. Results indicate high performance (average score ≥4.26/5), especially in step sufficiency and factual accuracy, suggesting that CoT reasoning by LLMs can support transparent and detailed clinical decision-making.

In Italy, the growing enthusiasm for artificial intelligence (AI) in healthcare contrasts with significant infrastructural, cultural, and trust-related barriers hindering its real-world adoption. Moving beyond the hype requires a systems thinking approach, proposing the learning health system (LHS) framework as a structured path for integration. We highlight the complementary roles of AI models: traditional machine learning (ML) is proven for diagnostics and prognostics, while large language models (LLMs) excel at administrative tasks and can structure unstructured data to train robust ML tools. The LHS cycle reveals key challenges for Italy: moving from Practice-to-Data requires overcoming data fragmentation; from Data-to-Knowledge involves transforming data into insights while mitigating bias; and from Knowledge-to-Practice necessitates bridging the gap between evidence and clinical workflow by building trust and AI literacy. Ultimately, successful and equitable AI implementation depends on a holistic strategy combining infrastructure development, multidisciplinary collaboration, and robust governance to enhance the quality and sustainability of the national healthcare system.

The combined use of IoT and AI enables automatic and precise collection of operative times through BLE bracelets, improving efficiency compared to manual recording. Surgery-specific models, trained on real data, better predict procedure duration, optimizing management and resources in the operating room.