At Principa, we are passionate about new ideas and product development. 20% of our revenue is ploughed back into innovation. One of our key areas of focus is machine learning where we have built machine learning solutions in both collections and the customer acquisition space. While our focus is primarily on credit risk and customer engagement, we are always interested in how machine learning has gained traction in other industries.
In this blog, we explore the application of machine learning and, in particular, deep learning in the medical field.
What is Deep Learning?
Deep learning is a type of machine learning that utilises artificial neural networks to learn patterns and then apply them in a range of activities. In some cases, the process outperforms human ability. Deep learning has been particularly effective in the spheres of image recognition, board games, lip reading and self-driving to mention a few.
Bringing a new drug to market can cost pharmaceutical companies between $350million to $5billion. There is an extensive process at play and a high failure rate at different stages. Machine learning and its subset, deep learning, can play a role in assisting the process.
Some of the key areas are listed below:
- Early-stage discovery – machine learning and deep learning are now used for the initial screening of compounds for the prediction of success in compound activity and interaction.
- Enhanced manufacturing – ML can be used to help reduce the costs of reproducing a drug.
- Personalised medicine – ML and its subset DL can use drug and patient data to predict responses to treatment.
- Medical trial cohort selection – using unsupervised DL to help select medical cohorts with specific genetic traits could reduce the size and cost of trials.
A concern with this approach is that DL may offer solutions that lack “explainability” which is outlined by some to be of strong importance. Although it should be noted that many medical treatments today lack explainability (for example many drugs used in anaesthesiology).
Human Genome Mapping
Whilst the human genome was first “mapped” in the 1990s at a cost of $2.7billion, the price of mapping a genome has been falling making it increasingly likely that many of us will have our genome mapped out in the future. The mapping of the 3 billion base pairs in a human genome results in 100GB of information. While parts of the genome are well understood, much of it is not. Assessing correlations and deterministic relationships between genotypes and phenotypes and, in particular, combinations of genes and mutations to predict susceptibility to conditions, or reaction to medical treatments is where deep learning is starting to have and will continue to play a significant role.
Companies such as Insight Genomics are utilising deep learning in oncology to help map treatments to tumours with genetic profiles.
Medical Image Analysis
In imaging, deep learning has been applied extensively from saving whales and classifying plankton to self-driving cars and describing photos. In medicine where visual diagnostics play an important role, so medical imaging is important too. Here are some examples.
Deep learning has been used to diagnose skin cancers simply from images. In a study published in September 2019, deep neural networks performed as well as trained dermatologists after the former was trained on a database of just over 4000 images. This is quite encouraging, as in certain countries, dermatologists are rare, and this provides an opportunity to improve healthcare. Companies such as Skin Vision provides an app where users can submit their skin pictures to be automatically screened for melanoma.
Just as dermatologists are in short supply, so too are radiologists who spend their careers analysing a variety of imagery. A paper published in November 2018 showed how Berkley data scientists were able to produce deep learning machines that performed as well as radiologists in Alzheimer diagnoses using PET scans. Another study in 2018 from Stanford showed convolutional networks outperforming radiologists in pneumonia diagnoses from chest X-rays.
All of this is very encouraging, and as with many AI advances, we don’t see a replacement of skills, but rather an augmented approach where the algorithms will assist the experts in the diagnosis process.
Machine learning is currently being used to predict epidemics around the world. An example of this is malaria outbreaks utilising data on temperature, humidity, rainfall and region in India.
More recently, the rise of the opioid epidemic has also provided an opportunity for machine learning. In the US over 2 million are addicted to opioids, and it is estimated that 33,000 died of opioid abuse in 2015. Machine learning is being used to predict where high rates of addiction are likely and where patient relapse is likely (utilising a phone app).
As with many other industries, the medical industry has ample areas of opportunity for the application of machine learning, including deep learning. We look forward to following the progress here. We hope to benefit from the learnings in health-care and apply them to other areas such as customer engagement and credit.