How is ML and AI helping to alleviate the COVID-19 pandemic?
In the past few weeks since the outbreak of the Novel Coronavirus (COVID-19) disease, there has been a widespread disruption in all major sectors, severely impacting the human way of life. Governments around the world are working closely with local authorities to track, respond and mitigate the effects of this pandemic and healthcare officials are resorting to analytics and advanced computational tools to augment their efforts to mitigate further infection.
In conjunction with all the great planning and research already taking place, the ML community has identified a number of arenas in which AI can help accelerate solutions. Some of these solutions contribute towards the vaccine development process, hospital management to test patients faster, forecast the spread of the infection, to name a few. But in the race to discover a breakthrough, I believe there needs to be a greater importance set towards identifying robust models that can learn from relatively smaller datasets and rapidly adapt to change in distribution. Previously trained models may not hold water with the dynamics set by this target domain and may fail during the shift between training and production distribution.
There have been a large number of interesting applications of Machine Learning to combat Coronavirus, I have listed a couple that I’ve come across and found noteworthy. Due to being fairly recent works, treat them with a level of skepticism.
- Prediction of criticality in patients with severe COVID-19 infection using three clinical features: a machine learning-based prognostic model with clinical data in Wuhan — The article helps distinguish the critical cases from the severe ones and hence to prioritise treatment between patients in a hospital to reduce the mortality rate. The model is based on the XGBoost algorithm to determine the probability of death based on certain features of the patient.
- Deep learning-based model for detecting 2019 novel coronavirus pneumonia on high-resolution computed tomography: a prospective study — The authors found that a great deal of time was being consumed by expert radiologist in reading the CT scans to detect the presence of COVID-19 pneumonia infected lesions. They employed UNet++ to develop the model using data from 106 patients and over 46,000 CT scans. The authors claim that it performed well in terms of precision and recall, has reduced the reading time of radiologist by 65% and for 27 patients, the model achieves comparable performance to an expert radiologist. Although, it must be noted that the data was accumulated from a single hospital and the performance of the model over a larger, more diverse dataset is still unknown.
3. Using Kalman Filter to Predict Corona Virus Spread — The article provides a promising method to predict the spread for a given infected region using the Kalman Filter algorithm on time series data. Although it predicts only a day ahead, this could prove vital in regions where the rate of infection growth is on a positive gradient.
4. Deep Learning Based Drug Screening for Novel Coronavirus 2019-nCov — The paper seeks to predict the effects of certain chemical compounds on the RNA sequences of Coronavirus (acquired from the GISAID database, not openly accessible and requires a .edu address). The model employed for this purpose is developed using a modified DenseNet, where the fully connected layer replaces the convolution layer, and predicts/ranks the protein-ligand interactions. Compounds such as Adenosine, Vidarabine, Mannitol, and a few others seem to return promising results.
5. A review of influenza detection and prediction through social networking sites — Based on the largely available unstructured data on social networking sites such as twitter, relevant information can be mined to effective predict data relevant to the spread and symptoms. Although this paper is not strictly focused on tracking the Coronavirus exclusively, it could be employed in situations where relying on hospital data is simply not enough. It lists out certain techniques that can be used to track the outbreak with the help of real time data from social media websites and blog posts.
There are a plethora of potentially impactful applications of ML out there to tackle the Coronavirus pandemic, however most of them are in the early stages of discovery. Some of the works haven’t been peer-reviewed or verified due to how recent the issue has been. Also, it is strongly advised to consult with an expert in the field (an epidemiologist, biologist, virologist, etc) before making assumptions on our study or even deploying to production. This acts as a layer of sanity to make sure our understanding — and therefore the understanding of the machine — of the problem at hand is not misinterpreted.
I believe there is still potential to be researched upon in terms of the ML techniques we can leverage towards uncovering a breakthrough of some sorts. One such method being the Continuous Domain Adaptation, where a neural network trained in a certain source domain can adapt to a target domain with minimal training data. Although this method has been previous applied in self-driving cars (to distinguish between the same obstacles under varying weather conditions), the same could be exploited in medical imaging (check this paper out — Towards Continuous Domain adaptation for Healthcare) where image segmentation of lung CT scans from a coronavirus infected patient could act as the target domain against a source domain trained from a previous pandemic such as SARS or MERS. This is especially useful since there are a limited number of datasets to train a whole new robust model from.
To sum up, ML continues to be an effective accelerant for data driven solutions and this epidemic has provided an opportunity for the community to step in and deliver. In conformity with the nature of AI, over time we’d have a better understanding of the problem at hand and help develop tools that will play an important role in containing and responding to both Coronavirus and any future pandemics that may come our way.
Until then, be smart and stay safe!
