Trustworthy AI tools for personalized oncology

Descripción del proyecto

Herramientas de inteligencia artificial para unos tratamientos antineoplásicos personalizados

El aprendizaje automático puede favorecer la medicina personalizada, ofreciendo una precisión y rapidez superiores a las del ser humano en el diagnóstico, la predicción de resultados y las recomendaciones de tratamiento. Sin embargo, garantizar la fiabilidad de las predicciones de un modelo a lo largo de su ciclo de vida es fundamental. El proyecto TAIPO, financiado con fondos europeos, tiene por objeto desarrollar herramientas de inteligencia artificial (IA) fiables para la oncología personalizada. Ello supone crear algoritmos fiables para diagnosticar y estratificar a los enfermos de cáncer, así como establecer un marco para la modelización transparente de resultados clínicos personalizados. En el proyecto se mejorará la fiabilidad de los modelos de IA para tres aplicaciones clínicas, a saber: clasificación de lesiones cutáneas, modelización personalizada de resultados en la leucemia mieloide aguda y recomendaciones terapéuticas en el cáncer de mama metastatizante. Se prevé que los resultados de TAIPO mejoren el aprendizaje automático de confianza en el campo de la medicina.

Objetivo

Modern machine learning algorithms have the potential to accelerate personalized medicine in a fast pace. To date, first tasks in medicine are being addressed with machine learning algorithms that surpass humans in terms of accuracy and speed, including diagnosis, outcome prediction and treatment recommendation. However, for a widespread adoption in clinical practice, a good performance in terms of speed and accuracy is not sufficient: practitioners also need to be able to trust a models prediction in all stages of its life cycle.
I will facilitate an efficient interaction of clinicians with AI models by developing trustworthy AI tools for personalized oncology: First, I will develop trustworthy AI tools and algorithms for diagnosis and stratification of cancer patients. Second, I will establish a framework for reliable and transparent modelling of personalized outcomes and therapy decisions in oncology.
TAIPO will result in novel algorithms and software tools for quantifying and improving the trustworthiness of AI models that I will apply to three clinical applications: (i) trustworthy AI-based skin lesion classification based on dermoscopic images, (ii) stratification and personalized outcome modelling for patients with acute myeloid leukaemia (AML) based on omics data, and (iv) therapy recommendation for metastatic breast cancer patients based on electronic health records.
TAIPO will increase the throughput of trustworthy diagnoses of skin lesions and pave the way for low-cost access to diagnostic care. It will empower clinicians to make personalized and reliable therapy decisions, which we will demonstrate at the example of AML and metastatic breast cancer. Our novel algorithms to evaluate and improve the reliability of AI models are a crucial contribution to close the gap between in-silico AI-bench and bedside and will further push the field of trustworthy machine learning with many applications of AI in medicine.