Artificial intelligence in healthcare

showcases the remarkable capability of AlphaFold in predicting protein structures, illustrating a strong correlation between the accuracy of the protein's backbone structure and its side chains. This correlation is crucial because it demonstrates AlphaFold's effectiveness in predicting the complex and detailed three-dimensional shapes of proteins, which are essential for understanding their functions in biological systems.

The article presents the development of PLIP, a multimodal artificial intelligence model that understands both images and text, trained on OpenPath, a large dataset of 208,414 pathology images paired with natural language descriptions. PLIP excels at classifying new pathology images across various datasets, achieving significant improvements in F1 scores compared to previous models. It also enhances pathology image retrieval from text inputs, showing substantial performance improvements across multiple datasets. One key performance metric of PLIP is its ability in zero-shot classification, where it achieves F1 scores ranging from 0.565 to 0.832 across four external validation datasets, outperforming the baseline CLIP model and demonstrating significant improvement in classifying pathology images.

In summary, this study introduces a novel method of using wearable technology and AI to track and predict the progression of Duchenne muscular dystrophy, offering significant improvements over current methods used in clinical trials and treatment development.

6-Minute Walk Distance

The paper is a scoping review evaluating the current landscape of randomized controlled trials (RCTs) for artificial intelligence (AI) algorithms used in clinical practice. The authors conducted a systematic search for RCTs published between January 2018 - August 2023, resulting in 84 unique trials included in the analysis. The key findings were:

• Most trials were in gastroenterology (41.7%), radiology (15.5%), surgery (6%), and cardiology (6%).
• The US (31%) and China (28.6%) had the most trials, with most being single-center studies.
• 54.8% of trials had primary endpoints related to diagnostic accuracy. Other endpoints assessed care management, patient outcomes, and clinical decision making.
• 82.1% of trials reported positive results for their primary endpoint, suggesting AI’s potential to improve healthcare. However, publication bias is likely.
• Most models were deep learning systems for medical imaging, particularly video analysis. Models for structured data used various techniques like neural networks.
• Industry developed most models (56%), followed by academia (39.3%). Models were commonly evaluated in an assistive setup with clinicians.

Page 2: Key implications and limitations:

• Need for more multi-center and international trials to ensure generalizability of models across diverse populations and settings.
• Focus has been on technical performance, but impact on patient outcomes needs more assessment through endpoints like survival, symptoms, need for treatment.
• High reported success rate indicates promise of AI but is likely inflated by publication bias and lack of comprehensive, multi-center studies.
• Homogeneity of studies/models indicates immaturity of the field. More variety needed in research groups, interventions, outcomes assessed.
• Understanding of AI’s limitations requires focus on unsuccessful trials and models.

In conclusion, the review reveals expanding interest in clinical AI trials across many fields and geographies. While most find positive results, publication bias is likely. More comprehensive research is essential to fully understand AI’s impact and limitations in healthcare. Key gaps include multi-center trials, clinically meaningful endpoints, and unsuccessful models.