A young and growing discipline
The use of Machine Learning applied to medicine took off in the 1990s but few models have found application in a medical context. Practitioners didn’t see the use, the models were difficult to train with little data, and generalization of a model with new data was rare.
Since the end of 2010, the discipline has entered a new era. What are the reasons for this expansion? The explosion in the amount of medical data, the standardization of this data and the recent advances in Machine Learning.
Explosion in the number of medical data
Driven by the policies implemented by the Obama administration, the adoption of the EHR (Electronic Health Records) in the USA has increased nine-fold since 2008 (9.4% in 2018 compared with 83.8% in 2015). This generation of data has enabled the creation of data sets needed to train Maching Learning algorithms.
- MIMIC was created in 2016 by MIT (open source). MIMIC is an open source database developed by MIT, including “non-identity-based” health data generated by approximately 60,000 ICU admissions. It includes demographic data, vital signs data, laboratory test results, drug prescriptions, etc.
- As part of the Precision Medicine Initiative, the Obama administration has launched the creation of a giant database. More than one million volunteers are going to share their health data: tests, EHRs, prescriptions and test results.
- Data sources have also multiplied in 2010 with the extension of Quantified Self (connected watches, health applications available on cell phones). The growth of disciplines such as genomics and proteomics have also contributed to the explosion of health data.
Standardization
In parallel with the creation of these datasets, which are essential for researchers, work has been carried out in recent years in the USA to standardize the data. Standardized data are available to researchers via APIs and allow researchers to standardize their work and have more qualitative data.
Some examples of standards in the field of diagnostics and drugs:
-ICD-9, ICD-10, ICD-11 (International Classification of Diseases)
-LOINC (Codes for laboratory examinations)
-NDCs-National Drugs code (Code for drug prescriptions)
-UMLS-Unified Medical Language system.
The revival of the Machine Learning since the end of the 2000s
This renewal is based on the advent of Big Data and the development of more powerful algorithms (Back-Propagation, CNN, RNN, Q-Learning). But also on the democratization of Machine Learning thanks to the work of evangelists on YouTube, the availability of languages dedicated to Machine Learning (Tensorflow, Keras, PyTorch) and the possibility of training models on the Cloud (Azure Machine Learning, Amazon’s AWS SageMaker, or Google Cloud AI Platform, Google Colab open source).
An attractive market for start-ups and GAFAs
The interest in Machine Learning applications in the healthcare field is illustrated the growth of investments :
- Digital Health Funding in 2011: $1.2Bby
- Digital Heath Funding in 2018: $6.9B
The number of start-ups in this field has also exploded. Tech giants have entered this field of activity: Google Health (which has integrated DeepMind Health), IBM Watson Health, Apple Health, Amazon Care. These large multinationals are buying up the most promising start-ups such as Truven Health Analytics, which was acquired by IBM in 2016 for $2.6 billion.
The Machine Learning applied to health is unique
The Machine Learning applied to healthcare is unique because of the diversity of stakeholders, data sources, fields of application and challenges.
Diversity of stakeholders
All actors in medicine: caregivers, patients, insurance companies, laboratories, research are likely to benefit from the Machine Learning applied to medicine.
Numerous fields of application
- Early diagnosis of serious and rare diseases.
- Daily health monitoring (preventive medicine)
- Prediction of diagnosis and care.
- Creation of new molecules and drugs.
- Better targeting of clinical trials.
- Prediction of the evolution of chronic diseases over time (treatment A versus treatment B).
Numerous challenges to take up
- Improving the robustness of algorithms: the algorithms developed have a power of life or death. They need to be extremely robust.
- Avoid demographic and social biases: the models should avoid inequalities in the healthcare system.
- Obtain qualitative data: it is often difficult to obtain data from insurance companies (access to paid data). It is also necessary to make data anonymous to respect privacy and medical confidentiality. A lot of data is missing because the interlocutors who collect the data are numerous (general practitioners, health insurance, health insurance, laboratories, etc..).
- Generalize models with new data: creating models adapted to all countries and all hospitals remains extremely difficult because of the diversity of medical equipment, data collected and work processes.
- Introduce causality in models applied to healthcare: the performance of care and the evolution of the disease depends on several factors.
This article was written from these resources…
