Jan 17 news - Further information
Report: Re-work deep learning workshop in healthcare, London 28/1 and 1/3 2017
Making sense of the overwhelming quantity of medical data generated everyday in clinics and research is still an ongoing effort, one that has seen an incredible acceleration in the last few years, also due to breakthroughs in the (apparently unrelated) fields of artificial intelligence (AI) and statistical learning, often also called machine learning. The "Deep Learning applications in Healthcare" workshop, wants to explore this intersection: how AI, data science and personalized medicine will shape the future of healthcare, what is the current state of the art and what are the issues yet to overcome.
The topics discussed ranged from technical details about the inner workings of the models, to discussions about policies, data ownership, privacy rights, ethical and legal concerns, to how cutting-edge medical technologies are putting the patient at the centre of the diagnostic process.
But, before diving in, a short primer on Deep Learning is needed:
Fundamentally, Deep Learning (DL) is a branch of Machine Learning (ML), and is based on a set of algorithms that, often with the use of so-called artificial neural networks, attempt to model high-level abstractions in data.
The architecture of these artificial networks is inspired by how our brain processes stimuli. In a simple example, there might be two layers of fully interconnected elements, called neurons: When the input layer receives a signal it passes on a transformed version of the signal to the next layer. In a deep network, there are many processing layers, up to an hundred, performing multiple linear and non-linear transformations between the input and output.
During the training process the network "learns", from thousands of examples, the right order of transformations that associate an input to its correct output. But, most importantly, it is then able to evaluate previously unseen inputs and perform certain tasks, like a classification: how to distinguish a cat from a dog in a youtube video, or an healthy cell from a cancerous cell in a histological sample.
These techniques have been successfully applied to fields like computer vision, automatic speech recognition, natural language processing, audio recognition and of course also in bioinformatics and biomedical image analysis, where they have been shown to produce state-of-the-art results in various tasks.
Deep Learning also made it to mainstream news in 2016, thanks to AlphaGo, an algorithm that could beat the best human players at Go, an ancient Chinese board game considered to be much more complex than chess.
But many more applications in healthcare are already available or are on their way.
Here is the (incomplete) list of the most interesting and exciting:
- Oladimeji Farri, Philips Research:
"Deep Learning-based Diagnostic Inferencing and Clinical Paraphrasing"
Starting from a text description of the symptoms of a patient, the model, trained using a medical knowledge-base of 6000 Wikipedia articles from Clinical Medicine Category, could give as output the probable diagnosis and possible pharmacological treatments.
- Polina Mamoshina, Insilico Medicine:
"Application of Deep Neural Networks to Biomarker Development"
- A Deep Learning-based method could predict the age of a person ( +/- 5y) from 40 commonly measured blood biomarkers
- A Deep Neural Network trained with examples of thousand of known anti-cancer molecules could predict novel chemical structures for anti-cancer drugs.
- Johanna Ernst, University of Oxford:
"10,000 Steps; So What? Are Wearable Technologies the Future of Clinical Trials?"
How to make sense of accelerometer data in clinical trials using Deep Learning, going beyond "step" counting as marker for physical activity
- Bashar Awwad Shiekh Hasan, University of Newcastle:
"Micro EMG: Imaging the Inner Structure of the Human Muscle"
Following the development of a new multi-channel Electromyography needle, the speaker presented a deep learning signal deconvolution method that allows to localize over 50 muscle fibres simultaneously, within 100 micro meter of accuracy.
- Fangde Liu, Imperial College London:
"SurgicalAI: Can we make Surgeries Autonomous?"
Deep Learning-based autonomous surgical planning system, based on computer vision
- Daniel Nathrath, Ada Health:
"The AI Will See You Now: Will Your Doctor Be Replaced by an Algorithm?"
Deep Learning-based iPhone application for medical consulting: a chat-bot that can ask information about symptoms, give medical advices and contact a physician when necessary.
On these premises, the health sector seems to be one of the most promising fields for Deep Learning-based applications: the availability of relatively cheap high-throughput techniques like DNA sequencing, the advent of wearables and the Internet of Things, Electronic Medical Records; with all this treasure of data available, and such powerful tools, building ML/DL applications for healthcare shouldn't be much of a problem.
Except that, for a number of reasons, we are not quite there yet:
First issue: how to share and use more medical data with more stakeholders, for more purposes, all while ensuring data integrity and protecting patient’s privacy. Today humans must manually manage medical data among clinics, hospitals, labs, pharmacies, and insurance companies. It is not working well, because there is no single list of all the places where data can be found or the order in which it was entered. Add to this the existence of too many data formats, many of which are proprietary (i.e. patented/private, not freely accessible).
Another big issue is that at the moment, to be reasonably accurate, Deep Networks need an enormous amount of data and expensive hardware to store the data and train the models.
And to be useful, data has to be prepared and curated by an human expert, that has to manually annotate or label the data, and this is a tedious and time-consuming task.
Last but not least: physicians often do not trust these systems and their predictions.
After all, one of the major drawbacks of Deep Networks is that, despite their excellent predictive capabilities, it is sometimes almost impossible to explain "why" they gave such a prediction.
To build doctor's trust and profit more from Deep Learning, these models really have to become more interpretable. On the other hand, it will become almost indispensable for the new doctors to have some degree of training in data science.
Ageing is a complex process, one that concerns the human body at all scales, from the molecules inside our cells to our muscles, bones and brains: how can we make sense of this complexity? Deep Learning capacity to abstract and find structures in noisy data seems a perfect fit for this problem.
A possible application would be using Deep Networks to aggregate known measurements and biomarkers correlated with different aspects of ageing (metabolism, physical functions, brain functions) to create a comprehensive "score" that reflects the overall subject's health. This could be used to monitor the conditions of the patient over time and to assess the effect of a treatment (e.g. nutritional intervention, physical activity regime or pharmacological treatment).
Plenty of other ways of applying the power of Deep Learning to healthcare are being developed right now, in universities and companies all over the world.
Is it just hype or will it bring a real transformation in how we take care of ourselves and of our sick and elderly people? Will ‘big data’ and AI empower the patients and help doctors in their decision-making processes, or is it a stepping stone to the industrialization and monetization of medical care? We can’t tell yet, what is certain is that a healthcare revolution is on the way.