During the Deep Learning in Healthcare Summit in London last week we hosted the ‘Shaping Tomorrow’ startup session to showcase innovative startups applying cutting-edge deep learning algorithms and tools to advance healthcare and medicine. Daria Danilina, an MBA student from London Business School, attended the event and kindly summarised the startup presentations.
Applications for business: Medical diagnosis, identification of heavy metals in water, terrain classification for automobile industry
Key take-away: Humans are trained to identify certain patterns. However, we tend to overlook things which we do not expect to see or are not trained to detect. In addition to this, anomalies exist that are impossible to identify for human eyes, such as tumours composed of soft tissue. Deep learning can help doctors identify such anomalies by training on images of a “healthy” human body and detecting anything that falls outside of the norm. Outside of healthcare, Cosmonio’s algorithms are deployed by a water supply company to detect bacteria in fluorescent images. The system can be easily deployed by new clients using a two phase approach. During phase 1 the neural network is trained using customer data and during phase 2 a domain specialist spends a couple of days improving the architecture. As soon as the system gets deployed, it gets better, as the algorithm never stops learning.
Key take-away: Implicit knowledge is a crucial component of the Data - Information - Knowledge - Wisdom framework. Domain experts spend a long time acquiring that knowledge. However, when they go on holiday or retire, that knowledge is lost. Capturing learning into a model is the promise of deep learning. Data differ in terms of its attributes (volume, velocity, variety, compliance and availability) and one person produces different data in different functions. In order to apply deep learning in a healthcare related context, we need to change the flow of information and continuously iterate and develop the algorithm. Labelling needs to be done for unsupervised learning and training can take up to two weeks. In addition to this, the production environment and big data stores pose an engineering challenge. If these challenges are conquered, a deep learning system can be applied in a variety of situations.
Detecting Dementia Before it Happens
Technology advancements: Structuring of data in a way that allows to trains an algorithm to predict when a patient with mild cognitive impairment will develop dementia
Applications for business: Identifying which patients will start developing dementia will make clinical trials of drugs targeting dementia more efficient and potentially help develop a system for managing the symptoms of the disease
Key take-away: Dementia patients account £26bn annual spending in the UK alone, with every third person in the UK expected to be affected. The earliest stage of the development of the disease is mild cognitive impairment. However, only 15% of patients displaying the symptoms progress developing dementia every year. Predicting which patients are going to progress from mild cognitive impairment to dementia can save time and money during the process of clinical trials for dementia-related drugs. Similar to other start-ups, the main challenge facing Avalon AI is the availability of data to train the algorithm. By getting access to 70,000 brain scans the team was able to develop a system that can predict the time until a patient will develop dementia as well as the biological age of the brain.
MedicSen Artificial Pancreas: Deep Learning to Improve Life Science Outcomes
Technology advancements: Cloud based neural network can give personalised treatment to diabetes patients
Applications for business: An integrated system that takes the data from a sensor, which a diabetes patient is wearing at all times, and uses these data to continuously monitor the well-being of a patient and predict when an injection will be needed
Key take-away: With $700bn annual spending Diabetes is not only a huge cost to the healthcare system but also a big social problem. Patients need continuous monitoring to be able to administer insulin injection at the right time. Despite a lot of research being done, precision and accuracy remain a challenge. Deep learning is capable of solving this problem but needs more data to be effective. Future areas that be tackled with this method include screening and diagnosis in emergency rooms, pharmaceutical studies , chronic and rare diseases. One key challenge is to maintain the relationship between doctors and patients. Another challenge is the lack of clarity around data laws and data security as well as liability of different parties and knowledge about neural networks. A multidisciplinary team is needed to tackle these challenges and bring a product to market.
Optimising CRISPR Genome Editing Using Machine Learning
Technology advancements: DESKGEN operating system for users to design and execute genome engineering experiments from a computer
Applications for business: Desktop genetics helps researchers to manipulate genomes and enables individuals to do genome editing right from your computer. Their DESKGEN genome editing software platform allows researchers to expertly design and perform CRISPR-based genome editing experiments in virtually any cell line or species.
Key take-away: CRISPR is a programmable way to alter genes in precise locations that can be compared to a “find and replace” command in a text editor. Application of CRISPR is something scientists have been waiting for for 90 years. Two weeks ago, doctors were able to rip the HIV virus completely out a child’s genes thanks to CRISPR. In addition to this, CRISPR gives hope to leukaemia patients by reprogramming their white blood cells, which was previously considered impossible. Desktop Genetics was founded when CRISPR wasn’t enjoying a lot of success. Unlike image recognition, deep learning in a biological context poses particular challenges related to data that needs to be acquired. Such data is expensive in terms of both time and cost, which means that algorithms have to work with noisy data sets. Dimensions of data get very big very quickly, posing a significant risk of overfitting. In addition to this, there is a lot of selection bias in the available data, whereby only those cases are published where the predictor worked. DESKGEN aims to make CRISPR gene editing more predictable, accessible and efficient, allowing scientists to scale their genomics research. The company’s dataset is up in Github and an API is will be released soon.
Are you a startup working on cutting-edge technology, or know of one that is? Suggest a startup for our next Shaping Tomorrow session via email to firstname.lastname@example.org