ML in Healthcare – Twelve real-world use cases Part3
Machine learning is exponentially increasing the speed of scientific research and discovery across domains, and the healthcare and medical space cannot be exempted, by any stretch of the imagination. Ranging from language process utilities that quicken-up research to predictive algorithms that caution healthcare personnel of an imminent heart attack, ML is complementary to the insights of a human agent and their practice cross healthcare disciplines.
But, with “solutionism” being all the rage with regards to AI, and ML technologies, medical providers are naturally wary with regards to how it will actually assist patients and provide a return on investment. Several AI solutions available on the market targeted at healthcare applications are customized to identify solutions to a really particular problem, like detection of the risk of having sepsis or diagnosis of an unknown strain of breast cancer. These innovative AI solutions make it tough or even not possible for enterprises to personalize their frameworks and models and see the highest ROI.
Open-source data science facilitates healthcare firms to take up models to tackle a plethora of hurdles leveraging the most advanced ML tech, like audio/visual processing. Leveraging open-source utilities, data scientists can customize applications in a fashion that fulfils healthcare’s IT’s stringent requirements and enhances patient care in a plethora of settings, ultimately distinguishing an enterprise from it’s competitors in the market. The following are 5 ML use cases within the medical space that can be produced with open-source data science tools and adapter for divergent functions.
1. Natural Language Processing (NLP) leveraged in admin tasks
A research carried out by the New England Journal of Medicine previous year identified 83% of respondents specified physician burnout as an issue within their business. 50% of them made the report that “off-loading administrative tasks” would be a beneficial solution to the issue, facilitating practitioners to spend additional time with their patients. A considerable proportion of these administrative tasks consists of review and updates of Electronic Health Records (EHR). Almost all hospitals in the United States leverage an EHR framework and so do a majority of the clinics.
Enhancing the effectiveness and efficiency of updating EHRs is a pressing priority for a lot of people. There is where NLP tools and utilities come into the picture, and, as one of the most innovative use cases, demonstrates the potency of AI/ML.
Through harnessing NLP tools that leverage algorithms to detect and categorize words and phrases, physicians can go about dictating notes straight to EHR during the course of patient visits. Practitioners and patients alike can undertake reviewing of charts and summarizations collated in an orderly manner by NLP tools rather than having to go through notes and test outcomes to comprehend a patient’s holistic health. Lesser the time expenditure with regards to the maintenance of EHRs, practitioners can get more hands-on and spend extra time with their patients.
2. Patient Risk Identification
Globally, healthcare providers have started harnessing utilities developed from ML models that leverage anomaly detection algorithms to forecast heart attacks, strokes, sepsis, and other grave medical complications. These utilities leverage data from patient’s historical data, everyday evaluations, and measurements of vital indicators in real-time, like heart rate and blood pressure, to caution staff with regards to impending patient risks so they can instantly take preventive actions.
One instance is El Camino Hospital. Their researchers and analysts leveraged EHR, bed alarm information, and nurse call information when a patient is susceptible for falling so they can take appropriate proactive, and preventive action to mitigate risk. There was an outcome of fall reduction by 39%. Going by the Joint Commission for Transforming Healthcare, an in-patient injury owing to a fall extends hospital stay by approximately 6.3 days and increases expenses by $14,000. Another instance is the Sepsis Sniffer Algorithm (SSA) put out by the Mayo Clinic.
The SSA leverages demographic information and vital sign measures to set off an alert whenever the risk of contracting sepsis escalates, reducing manual screening by 72%. This facilitates practitioners and nurses to spend extra time in treatment of the illnesses patients initially came complaining about it.
3. Accelerating Medical Research Insight
Analysts and healthcare practitioners would have to undertake an in-depth reading and process a massive quantity of reports and research to stay updated with trends in particular areas of healthcare research. For instance, academics put out more than >342,000 articles on drug assessment and analysis performed between the years 2007 and 2016. Leveraging NLP utilities and neural networks in parsing literature will furnish healthcare researchers with worthwhile insights in the years that lie ahead of us.
For instance, a unit of analysts and researchers from Ireland and the United States collaborated to carry out a research on Adverse Drug Events (ADEs) harnessing text mining, predictive analytics, and neural networks to undertake analysis of massive databases of medical literature and social media posts for comments that are made in relation to drug adverse effects. Upon analysis of 300,000+ articles from healthcare journals and more than 1.6 million responses on social media, the unit harnessed data visualization utilities to demonstrate the relationships amongst drugs and their side effects.
NLP is additionally being harnessed in the pursuit of mining unstructured data in EHRs for obtaining critical insight, like data from electrocardiogram results or copies of manually authored notes that were uploaded to a patient’s record, however, not inputted into form fields. cTAKES is one instance of an open-source NLP undertaking by the Mayo Clinic, Boston Children’s Hosptial, and other entities to generate a utility that parses unstructured data within EHRs to collect insights from the data.
4. Visual Data Processing for Tumor Detection
Radiologists workloads have escalated considerably in the previous few years. A few research articles have discovered that the average radiologist in engaged in interpretation of an image every 3-4 seconds, based on the demand that they command in the market. Researchers have generated deep learning algorithms that have undergone training on prior captured radiographic imagery to recognize the early developments of tumours within the lungs, breasts, brains, and other regions. Algorithms can undergo training to identify advanced patterns in radiographic image data.
They can identify breast cancer from mammograms with considerable precision. One such detection tool that snuffs out breast cancer in the preliminary stages has been produced by the Houston Methodist Research Institute; it undertakes interpretation of mammograms with 99% precision and furnishes diagnostic data 30 times quicker than a human. Utilities such as these, assist in enhancing patient care. They make them better performers at their jobs, however are not substitutes for them.
5. Leveraging Convolutional Neural Networks (CNNs) for skin cancer diagnosis
CNNs are potent utilities for identification and classification of imagery. Various researchers have harnessed them to produce ML-driven models for skin cancer identification with 87—95% precision leveraging TensorFlow, scikit-learn, keras, and other open-source utilities. By contrast, dermatologists have 65% to 85% precision rate in detection of melanomas. Models undergo training leveraging thousands of images of malignant and harmless skin lesions. One instance of an open-source project such as this one is available in the public domain, on websites such as Github. On top of skin cancer diagnosis, analysts are also leveraging CNNs to produce utilities for diagnosis of tuberculosis, heart illness, Alzheimer’s disease, and other conditions.
While healthcare entities ought to be more prudent than a majority of other industries with regards to security, compliance, and governance, they can still undertake training of ML models leveraging anonymized data in order to maintain compliance with HIPAA specifications. Guaranteeing the integrity of the software environment is critical for managing healthcare client data. The next part of this multi-part blog series will go into detail with regards to compliance. It also goes further into discussing how AI/ML assist in maintenance of healthcare compliance regulations.