In a typical 150 bedded government civil hospital we would usually see close to 85,000 patients in a month and conduct approximately 6,000 radiological services with ~ 50% of them being Chest X rays. Expecting radiologists to read and report each one of them is not a feasible solution especially when the country is facing a shortage of skilled healthcare professionals. We intend to discuss the specific use of notification-only triage workflow for Chest X-rays in a government hospital.
Artificial intelligence in radiology has undergone a massive transfiguration from a few innovative solutions to a plethora of platforms and algorithms, it has grown to be a technology might that can change healthcare delivery all by itself. Nevertheless, like majority of technology solutions AI is not close to perfection and also not a replacement for trained, skilled clinicians. It is a decorated collection of algorithms, deep learning, machine learning programs, complex neural networks that are keen to change how radiology would be delivered as a service.
Out of the 50+ FDA approved radiology AI Algorithms, few are noteworthy of showing signs of a greater impact on government healthcare in India.
Case in Point
Pneumothorax, a collapsed lung is a potentially life-threatening condition which can be treated effectively as an emergency. It can be detected by Chest X-ray as it presents radiological features such as a visible visceral pleural edge is seen as a very thin sharp white line, no lung markings are seen peripheral to this line, peripheral space is radiolucent compared to the adjacent lung. Expecting these features to be accurately identified and read by a non-radiologists is a taller ask.
Tuberculosis:
India has the highest TB burden in the world, having an estimated incidence of 26.9 lakh cases in 20192. The burden of undetected tuberculosis is large in many settings especially in high-risk groups. Accidental detection of “Diffuse miliary opacities” on an unintentional chest X-ray also have been found to of value. In real time Indian public health scenario, the likelihood of this being missed to be read and diagnosed by a general physician in a CHC is fairly on a higher side.
Imagine if we have a machine that detects Pneumothorax and Tuberculosis and alerts the relevant stakeholders in the care cycle.
Building a Narrative:
A 35-year-old patient is admitted in emergency room after a Road Traffic Accident (RTA), and X-ray Chest is done in the ER. An artificial algorithm reads the X-ray image and classifies the chest X-ray as pneumothorax present, absent, or unclear. In cases when the algorithm returns pneumothorax positive and a radiologist is not present at the time of imaging, an alert is provided to the treating physician. If a radiologist is present, the exam is prioritized in the radiologist worklist for urgent interpretation and reporting.
A 45-year-old patient walked into a CHC, complaining of chest pain and was advised a chest X-ray with a battery of other tests. An algorithm evaluates the image and categorizes the chest as being normal, abnormal and further states the extent of tuberculosis findings. In cases when the algorithm returns a positive TB and a radiologist is not present at the time of imaging; an alert is provided to the treating physician. If a radiologist is present, the exam is prioritized in the radiologist worklist for urgent interpretation and reporting.
Workflow
The image is acquired from the X-ray machine and sent to Picture Archival and Communication System(PACS) and AI system. The image is then read by the AI system which detects and estimates pneumothorax size or Tuberculosis manifestations. An alert message is sent to PACS from the engine with the information, identification and also highlighting the abnormality.
Making it Work
Though these AI is expected to change radiology services, education is necessary to teach individuals about AI tools and data science. This needs to be multidisciplinary and involve various healthcare leadership, clinical teams, AI experts and patients.
There would be no second thoughts on the impacts it can bring to the way we deliver care, however getting it through the system and institutionalizing an AI practice would be a greater challenge in Indian Public Health.
Unlike other IT projects implementing a radiology AI would require a basic understanding of the machine learning process and project management capabilities emboldened by patience and confidence.
The below mentioned Diagram depicts the classic steps of a machine learning process.
For a government organization in healthcare it becomes imperative to have all the 7 steps done in our own facilities to enable utmost ethical consideration. Algorithms might have been developed outside the ecosystem but testing it on our Datasets of Images becomes critical.
Data Collection and Data Preparation: -Data collection of X-ray images should not pose a challenge as DICOM images from the modality are available in PACS systems. Data preparation should be done in collaboration with pulmonologists and radiologists to ensure we add the normalized, outliers and images of varying complexities of the said diseases.
Choosing a model is the key to the success of the project, while we have various models of reading an X-ray it becomes imperative to do localized testing of these methods and choose the model like NLP which was selected by NIH for some of their AI programs. In case of estimating the percentage volume of a pneumothorax there are multiple methods like the Collins Method, Rhea Method and Light Index, we need to explore each of them and select the one apt for an AI to pick up.
During the initial evaluation phase, it is quite common and expected to see some results of the grid and getting it right during the parameter tuning phase should be the sole intention of this phase.
The results must be understood and analyzed by different perspectives statistical validity, considering the reproducibility with different cohorts and correctness of statistical values obtained (i.e., metrics) and intra-validity, regarding the clinical and real implications of the algorithms on a daily basis (i.e., clinical effectiveness). This is a pairwise co-existence; none of the ML algorithms will be applied in clinical routine if there is no agreement from both sides.
The validity of the algorithms, a whole imaging data set should be split into 3 different subgroups, called training set, validation set, and testing set, respectively. These groups are often selected in such way that subgroups share demographic distributions such as age or sex, in order to represent a real-world scenario. Accuracy measures the percentage of the algorithm classifying the input data correctly. One of the drawbacks of using accuracy as the metric is that there is a knowledge loss when measuring False Positive and False Negative observations. Therefore, Specificity (Sp) and Sensitivity (Se) are widely used for measuring the performance of ML in Image Analysis. An acceptable Sensitivity, Specificity and Predictability should be agreed upon and a roadmap to increase the same must be in place for the success of the project.
A rigorous time bound Pilot should be conducted before an enterprise roll-out. Once deployed these solutions of Notifications to Treating Physicians and Radiologists will not only be a Triaging solution but serves the needy. Detecting an illness amounts to saving a life which is the most important to clinicians and we just got an intelligent partner-Artificial Intelligence to augment us.
(Disclaimer: Dr. Anil is a Clinician by education and into Healthcare-IT.This views, thoughts and opinions expressed in this article are author’s own and does not represent or relate to the APAC Network)
References:
- Chest radiography in tuberculosis detection – summary of current WHO recommendations and guidance on programmatic approaches. World Health Organization. ISBN 978 92 4 151150 6.
- India TB Report 2020, NATIONAL TUBERCULOSIS ELIMINATION PROGRAMME, Central TB Division Ministry of Health and Family Welfare, Govt of India.
- Image-Based Cardiac Diagnosis With Machine Learning: A Review, Carlos Martin-Isla, Victor M. Campello, Cristian Izquierdo,ZahraRaisi-Estabragh,BettinaBaeßler, Steffen E. Petersen, and Karim Lekadir
