Title: MedPerf: Enhancing Medical AI for Clinical Efficacy
MLCommons, a global engineering community, has developed MedPerf, an open benchmarking platform, to evaluate AI models on real-world medical data. This platform aims to improve the clinical translation of Medical AI by eliminating bias and increasing generalizability while protecting patient privacy and minimizing legal concerns.
The Significance of Evaluating AI Models on Real-world Data:
Medical AI models trained on limited clinical data may unintentionally develop biases against certain patient populations. This can reduce their effectiveness in real-world settings. However, due to privacy and regulatory concerns, accessing diverse datasets for training is challenging. MedPerf addresses this issue by providing convenient and secure access to global medical data, allowing researchers to evaluate and validate AI models.
Streamlined Evaluation Process:
MedPerf enables healthcare organizations to evaluate AI models in a streamlined and human-supervised manner, even without access to patient data. The platform facilitates remote installation and on-premises review of medical AI models by data suppliers. This federated assessment approach alleviates concerns about patient privacy and builds trust among healthcare stakeholders.
Efficiency and Effectiveness:
MedPerf has proven its effectiveness in the largest federated experiment on glioblastoma, the FeTS Challenge. It facilitated the benchmarking of 41 distinct models across 32 sites on 6 continents, demonstrating the platform’s ability to evaluate numerous AI models with collaborators in hours rather than months.
Pilot Trials and Applications:
MedPerf has undergone pilot trials in several academic medical research areas. These trials have focused on brain tumor segmentation, pancreatic imaging, and surgical workflow analysis. The findings from these trials confirm that federated evaluation benchmarks contribute to more accessible AI-enabled medical care.
Wider Adoption and Support:
MedPerf promotes the use of widely used ML libraries like fast.ai for their usability, adaptability, and performance. The platform also supports various API-only and private AI models, including Microsoft Azure OpenAI Services, Epic Cognitive Computing, and HF inference points.
Expanding Use Cases:
Although initially designed for radiography, MedPerf can be applied to any field of biomedicine. Its sister project, GaNDLF, simplifies the construction of ML pipelines supporting digital pathology and omics activities. To bridge the data engineering gap, MedPerf provides access to state-of-the-art pre-trained CV and NLP models through specialized low-code libraries like PathML or SlideFlow, Spark NLP, and MONAI.
The MedPerf team aims to boost confidence in medical AI, accelerate the adoption of ML in clinical settings, and enable personalized care, cost reduction, and improved quality of life for doctors and patients.
With MedPerf, MLCommons has developed a powerful benchmarking platform that enhances medical AI’s clinical efficacy. By providing access to diverse real-world medical data in a secure and streamlined manner, MedPerf enables the evaluation and validation of AI models, contributing to the advancement of AI-enabled medical care.
– To read more about MedPerf, please visit the [Reference Article](https://mlcommons.org/en/news/medperf-nature-mi/).
– For additional information about the project, visit the [Project Page](https://www.medperf.org/).