Anything that can be automated should be automated - if you believe in that, healthcare is probably the industry where automation driven by artificial intelligence (AI) and machine learning (ML) would truly make an impact, in terms of time, cost and saving lives. Many approaches incorporate complex artificial neural network architectures and are collectively referred to as deep learning (DL) applications. DL has been shown as capable of representing and learning predictable relationships in many diverse forms of data and hold promise for transforming the future of healthcare. The areas where DL has achieved remarkable results so far are around image detection and generation, natural language processing, time series forecasting, and so on. There are a few DL applications in healthcare that caught my eyes in recent years including the use of DL on imaging data for skin cancer detection, disease prediction using high resolution time series data from biosensors, and predictive analytics using electronic health record (EHR) data, in particular when combined with multiomics data to facilitate precision medicine.
A group of Stanford scientists published a Nature article in 2017 on the use of deep convolutional neural networks (CNN) for two binary classification use cases keratinocyte carcinomas versus benign seborrheic keratoses and malignant melanomas versus benign nevi using 129,450 clinical images. The CNN achieves performance on par with all tested experts across both tasks, demonstrating its capability of classifying skin cancer with a level of competence comparable to dermatologists. The application of this technology, in combination with mobile devices, can potentially extend the reach of dermatologists outside of the clinic with a projection of 6.3 billion smartphone subscriptions by the year 2021.
"Anything that can be automated should be automated - if you believe in that, healthcare is probably the industry where automation driven by artificial intelligence (AI) and machine learning (ML) would truly make an impact, in terms of time, cost and saving lives"
Popular wearables, such as Fitbit and Apple Watch, generate trillions of unlabeled sensor data points per year, including rich signals like resting heart rate and heart rate variability, which have been shown to correlate with health conditions as diverse as diabetes, sleep apnea, atrial fibrillation, heart failure, sudden cardiac death, and irritable bowel syndrome. In 2018, Cardiogram published a paper named “DeepHeart,” which is a semi-supervised, multi-task long short-term memory (LSTM) on 57,675 person-weeks of data from off-the-shelf wearable heart rate sensors, showing high accuracy at detecting multiple medical conditions, including diabetes, high cholesterol, high blood pressure, and sleep apnea. In the same year, Jawbone Health has published an article that went a step further using a deep convolutional-recurrent neural network for the detection of atrial fibrillation on 180 hours of photoplethysmography (PPG) data sampled at 20Hz. It achieved state-of-the-art performance validated in the experimental subjects which constitutes a significant improvement on previous results utilising domain-specific feature engineering, such as heart rate extraction, and brings large-scale atrial fibrillation screenings within imminent reach.
A group of scientists at the Department of Genetics and Genomic Sciences, Icahn School of Medicine at Mount Sinai published an article in 2016, named “Deep Patient,” using a three-layer stack of denoising autoencoders to capture hierarchical regularities and dependencies in the aggregated EHRs of about 700,000 patients from the Mount Sinai data warehouse that facilitates clinical predictive modeling. They demonstrated that the deep learning approach outperforms those achieved using representations based on raw EHR data and alternative feature learning strategies in predictive outcomes of 78 diseases from diverse clinical domains and temporal windows.
Precision medicine is an active area of research which gives hope to personalized, in other words, tailored-made treatment for patients based on deep understanding of the biological mechanism underlying the disease. In a clinical setting, precision medicine is only really practised in few disease areas such as cancer, rare diseases and infectious diseases. Omics is a wide domain involving specialized and high-throughput biotechnological methods, instruments, and algorithms. These techniques are often used to measure and study complex biological systems and their interactions, such as genomics, transcriptomics, proteomics, epigenomics, metabolomics, phenomics, pharmacogenomics, to name a few. Common Omics technology applications in precision medicine include disease biomarker identification, drug target identification, disease progression measurements, and optimization of treatment strategies.
Combining Omics and EHR data for precision medicine applications pose considerable challenges and opportunities for DL, the former of which is due to many factors such as low signal to noise, analytical variance, high dimensionality and complex data integration requirements. Omics data do not readily conform to the deterministic assumptions underlying many mainstream DL implementations and require domain-specific approaches for dealing with biologically unrelated modes of variance. For this to truly work, it requires a multidisciplinary approach on large amounts of fully linked data, perhaps yet to be collected and annotated, in different stages of biological process across many disease areas which requires buy-ins from all parties involved including academics, healthcare payers and providers, biotechnology and pharmaceutical companies, policymakers and patients.