Artificial intelligence and health care: four challenges to overcome – Actu IA

Whether it’s using a chatbot to help diagnose symptoms in telemedicine or improving medical imaging and patient monitoring, healthcare professionals are increasingly using the support of artificial intelligence.

But as AI moves beyond research alone and gains traction in clinical care and drug discovery, there remain a number of critical hurdles that will determine the success or failure of its use in healthcare. health and life sciences.

To overcome them, we must design solutions that take them into account – ensuring that data sovereignty, ethics and patient privacy are considered and respected at every step of the process. Moreover, the emergence of new techniques such as synthetic data and federated learning provide answers to some of these issues.

1 . Focus on ethics and governance

The European GDPR regulation is a model for AI applied to the health sector, but it is necessary to go even further. AI systems cannot be thought of as black boxes and patients must imperatively be informed about how their data is stored and used.

How the AI ​​will be used and ensuring that the data it was trained on accurately reflects the diversity of the population will need to be considered for each use case of the algorithm. Additionally, ensuring that these algorithms have been trained and annotated accurately is paramount to avoid erroneous conclusions and ensure the correct diagnosis is made.

This approach is similar to the work done for autonomous vehicles, which is to ensure that the training data is correct and ensure the safety of the roads.

2. Data annotation

The availability of well annotated data, by trained experts, is an additional obstacle. Programs such as the London Medical Imaging Center and the AI ​​Center for Value Based Healthcare, PathLAKE, and the Industrial Center of Artificial Intelligence Research in Digital Diagnostics (iCAIRD), are seeing data volumes soar from a handful images to a million in the space of a year.

This work is evolving from a great exercise in technical engineering to a particularly laborious task of rapidly labeling clinical data with great precision. Thus, the same care taken in the diagnosis of a patient with cancer must be taken in the annotation of a pathological image before it is integrated into an algorithm dedicated to the search for care.

As data volumes grow, we need to be able to keep pace with data curation and annotation to ensure it is relevant and useful. Luckily, we are now witnessing the development of semi-automatic annotation techniques, which provide valuable assistance in this task.

3. Growing adoption of federated learning

In addition to leveraging AI for data collection and analysis, federated learning techniques will improve data sharing across departments, institutions, and businesses while respecting data sovereignty and regulations. matters of confidentiality.

Federated learning is a privacy-preserving technique that brings the AI ​​model to local data, trains the model in a distributed fashion, and aggregates all learning along the way. This way, no data is exchanged or leaves the healthcare facility. The only exchange that occurs is in the gradients of the model.

Programs such as the AI ​​Center for Value Based Healthcare use federated learning to build more robust AI models. Similar public and private partnerships can rely on an open-source platform to ensure that data remains private and does not leave the institution.

4. Create synthetic data

Synthetic data offers researchers the ability to create tools, models, and tasks by simulating real data, which, not being associated with patients, can be easily shared between research institutions while ensuring confidentiality.

The data retains the characteristics of the medical records on which it was trained, but these AI-generated records could be used to complement and balance the datasets to better represent the patient population and help eliminate bias .

For example, research institutes could use synthetic data to create digital diabetes patient medical records with characteristics similar to those of a real population. King’s College London is working to generate synthetic brain images using this approach to better understand the evolution of brain diseases such as Alzheimer’s disease, with the aim of improving diagnosis and treatment.

In order to drive wide adoption of AI in healthcare, collaboration and coordination between governments, industry and technology players is needed to address all these challenges and improve model accuracy, to best support the healthcare professionals when making diagnoses and treatment decisions. We hope these tools will begin to accelerate the overcoming of these challenges, allowing AI to enter the clinical path as quickly as possible.

We want to say thanks to the writer of this article for this amazing material

Artificial intelligence and health care: four challenges to overcome – Actu IA

Discover our social media accounts as well as the other related pages