³ÉÈË´óƬ

Medical imaging refers to the set of techniques that produce visual representations of a body’s internal structures, revealing the anatomy hidden by the skin and bones as well as the function of some organs or tissues.

Medical imaging can be carried out using a range of technologies, including: 

• X-ray/CT scan
• Ultrasound
• Magnetic Resonance Imaging (MRI) 
• Positron emission tomography (PET) scan
• Endoscopy
• Medical photography.

Once a medical image is captured, it is analysed by a medical imaging specialist (e.g. radiologist) who looks for signs of disease or abnormalities and diagnoses the medical condition. The interpretation of medical images is traditionally a manual process that is both time-consuming and susceptible to the subjective assessment of the specialist. Furthermore, the volume and complexity of medical imagery is increasing faster than the availability of human expertise to interpret it (especially in low resource settings such as remote and regional areas). 

Machine learning can be used to help medical imaging specialists to:

• Automate time-consuming measurements
• Reduce the implications of cognitive biases
• Reduce ‘retakes’, which can cause unnecessary exposure to radiation
• Track the progression of a disease 
• Monitor recovery following treatment
• Identify incidental findings (previously undiagnosed medical conditions that are discovered during evaluation for another medical condition).

For example, using computer vision and deep learning techniques, computers can be trained to look for specific physical features that may indicate disease (such as cancerous tumours). 3D reconstruction techniques can be used to automatically generate 3D models from 2D scans, while machine learning techniques can be used to identify physical features from the image (e.g. bladder wall) and automatically perform time-consuming measurements. 

AIML researchers have developed an autonomous medical image program using machine learning techniques to detect breast tumours in MRI images. The unique approach used is 1.78 times faster in finding a lesion than existing methods and just as accurate. 

AIML has worked with medical equipment manufacturers, including , to develop machine learning techniques to :

• Measure bladder volume from ultrasound scans
• Measure heart health characteristics from ultrasound scans
• Perform retinal mapping
• Monitor the management and care of ulcerative wounds.

Pathology is a medical speciality that determines the cause of diseases by examining and testing: 

• Body tissues (from biopsies and pap smears) 
• Bodily fluids (from samples including swab, blood and urine). 

Pathology covers an extensive range of laboratory-based testing that is performed by specialist scientists and technicians. The results from pathology tests help doctors:

• Diagnose and treat patients correctly 
• Monitor a patient’s response to treatment and recovery.

Machine learning can be used to automate laboratory testing and data processing activities across the range of pathology activities, freeing-up pathologists’ time and enabling faster delivery of patient results.

As an example, in the pathological field of microbiology, analysts identify disease-causing microorganisms (bacteria, viruses, fungi and parasites) and perform antibiotic susceptibility testing. To identify the specific disease-causing microbes, analysts grow and identify colonies on a growth-medium (such as agar). Each of these growth-medium samples has to be examined to identify the presence of clinically significant microbial colonies. A busy microbiology laboratory may process thousands of growth-medium samples every day.

Microbiologists also undertake laboratory-based testing to determine the effectiveness of antimicrobial treatments against the specific disease-causing microorganism that they have identified. In Figure 2 the antibiotics are contained in the white paper discs. The effectiveness of the antibiotic is determined by measuring the radius of the circle that is clear of microbial colonies around the antibiotic disc.

Using machine learning it is possible to train a system to identify features that signify significant microbial colony growth on the growth media.

In collaboration with , AIML has developed machine learning techniques for use in the FDA-approved Automated Plate Assessment System (APAS) system. APAS is able to examine growth-media plates to detect the presence of significant microbial growth. APAS streamlines the plate triaging stage of the microbiology workflow, delivering reliable, consistent plate results and overcoming a major bottleneck in microbiology laboratory workflows (/aiml/news/list/2018/10/15/speeding-disease-diagnosis).

Screening tests are medical tests that aim to identify the disease before symptoms appear. The goal of screening is to detect disease at an earlier and more treatable stage. Screening tests may include:

• Laboratory tests to check blood and other fluids
• Genetic tests that look for inherited genetic markers linked to disease
• Imaging tests that look for irregularities that may indicate disease. 

Screening tests are typically offered to individuals in an eligible demographic group; usually defined by age, sex and other risk factors such as family history. In order for a screening test to be successful, the test must be simple and minimally invasive for the patient. 

Population screening tests in Australia include:

• Prenatal screening via maternal blood tests
• Newborn screening (commonly referred to as the ‘heel prick test’)
• Bowel cancer screening
• Cervical cancer screening
• Breast cancer screening.

The physical output from screening tests is a large number of samples and images that need to be analysed by pathologists and medical imaging specialists. Much of this testing is currently manual in nature and time consuming; with the analytical outcome being susceptible to the subjective assessment of the specialist. 

Machine learning techniques (specifically computer vision, 3D reconstruction and deep learning techniques) have been demonstrated to be effective at detecting pathologic changes, that are indicative of disease, at an early stage. Machine learning techniques can be used to automate laboratory testing and data processing activities. This, in turn, provides for better management of laboratory workflows, the freeing-up specialists’ time, the reduction of cognitive biases and faster delivery of patient results. 

The healthcare industry is evolving rapidly. The digitisation of medical records and the increasing availability of internet-connected health monitoring technologies mean that large volumes of data are becoming available to healthcare professionals. While this data potentially allows for faster and more accurate diagnoses, this is only possible if the large amounts of data can be analysed in a timely and accurate manner for incorporation into patient care. We are rapidly approaching the point where the volume, diversity and complexity of medical data exceeds a medical professional’s ability to interpret the data and recognise more complex patterns. 

Machine learning techniques can be used to collate and process the medical data that is available to medical professionals, highlighting clinically-relevant information in the huge quantities of data. Machine learning techniques can be used to: 

• Compile and process medical data to provide better traceability and faster, more consistent access to medical records
• Assist clinical decision making by identifying patterns and anomalies hidden in the data
• Cluster patients’ traits and predict disease outcomes
• Develop individualised treatment plans
• Monitor and track patients’ response to treatment
• Improve note-taking by voice using natural language processing 
• Extract information from unstructured data, such as clinical notes and medical journals to supplement and enrich structured medical data
• Highlight treatments mistakes, reduce workflow inefficiencies and help area-healthcare systems avoid unnecessary patient hospitalisations.

There are a wide range of conditions that are known to benefit from vital sign monitoring and routine collection of health survey data. This includes (but is not limited to) asthma, diabetes, chronic obstructive pulmonary disease, heart failure, coronary artery disease, cardiac arrhythmias, mental health, palliative care, home dialysis and wound care.

Remote Patient Monitoring (RPM) refers to the use of specific medical technology to enable monitoring of patients outside of conventional clinical settings; such as in the home. Incorporating RPM in chronic disease management can provide a more holistic view of a patient’s health over time, increase visibility into a patient’s adherence to treatment and enable timely intervention before a costly care episode occurs. RPM allows patients with chronic health conditions, the elderly and disabled individuals to live at home longer; avoiding hospitalisation or having to move into skilled nursing facilities.

RPM can employ a host of wearable, implantable and home deployable internet-connected measurement devices to collect a large range of health data, including vital signs, weight, blood pressure, blood sugar, blood oxygen levels, heart rate, electrocardiograms and movement/activity.

Another important and growing source of large quantities of remotely-collected health data are consumer-level Wearable Health Devices (WHD) (e.g. Apple watch, Fitbit and Garmin devices). WHDs continuously collect health data during different daily activities over days/weeks/months. WHDs are helping people to better monitor their health status (both at an activity/fitness level and at a medical level) and provide data to clinicians that may allow earlier diagnosis and treatment of certain medical conditions. WHDs routinely collect data on:

• Heart rate
• Activity levels
• Calories burnt
• Steps
• Location
• Sleep cycles.

Machine learning techniques can be applied to RPM and WHD data to:

• Collate and track data
• Identify patterns and derive correlations between cause and effect to aid in diagnosis
• Predict that a health event is imminent 
• Classify the severity of the health event
• Identify significant events/anomalies in the data that may indicate that medical intervention is warranted (e.g. high heart rate, heart arrhythmia or hard falls) 
• Automatically notify care providers and/or emergency responders.

In a medical emergency, the provision of timely and appropriate emergency medical care is fundamentally important in reducing patient mortality and morbidity.

In South Australia, 000 calls are answered by a human operator who asks whether the caller needs fire, police or ambulance assistance and then directs the call to the appropriate service. For example, if the person asks for ambulance, the call is directed to the SA Ambulance Emergency Operations Centre where the call is answered by the Emergency Medical Dispatch Support Officer (EMDSO). The EMDSO will ask for basic information, such as the name of the person with whom they are speaking, the name of the patient, the patient’s condition, the address and the caller’s phone number. The EMDSO assesses the situation and determines the priority of the medical emergency. 

When there is a large volume of calls, incoming calls are placed on hold until there is an operator available to answer the call. This can introduce a delay into the medical response, increasing the risk of mortality or morbidity to the patient. When the volume of calls gets too large for the system to handle, it has been known for the hold-queuing system to lose calls, or for the system itself to crash. 

Machine learning, specifically natural language processing, can be used to power virtual assistants that have the capability to: 

• Answer and direct multiple calls in parallel
• Take basic call information 
• Highlight high-priority calls 
• Provide multilingual services
• Incorporate vocal-stress analysis 
• Analyse non-vocal queues, such as traffic noise or breathing. 

Once an ambulance has been dispatched, it is important that the ambulance gets to the patient as quickly as possible. Machine learning can be used to collate data such as traffic, road conditions and the weather to find the fastest route to the patient and then to a hospital. 

Remote monitoring of the patient from the ambulance can provide the hospital with important information on the patient’s condition. Machine learning can be used to collate the patient’s medical data to streamline the hospital triage process and allows the hospital to prepare for the patient’s arrival while the ambulance is en-route.

Contact AIML for more information on how machine learning can streamline your emergency response activity.