By Harit Phowatthanasathian
With over eight thousand human diseases and the list growing, our understanding of medical conditions currently does not match its vastness. Around five percent of first admission patient conditions are misdiagnosed due to the overlapping symptoms, lacking experience of healthcare workers, or inadequate protocols (How Common is Misdiagnosis – Infographic, 2020). This begs the question of how humanity can continue to elevate the standard of healthcare?
Machine learning and deep learning algorithms have appeared along with the explosive advancement of computational power. With the new niche of computational biology, the implementation of machine learning into the healthcare system has been a growing priority. The central premise of machine learning in healthcare is to utilize the growing availability of medical databases to train prediction programs that will effectively aid as a second opinion or supplementary advice for medical professionals. Technology giants, including Google and Microsoft, have invested their computational power and time into developing these algorithms with repetitive, protocol-driven tasks, including reading MRI imaging, analyzing mammograms, and identifying skin cancer from collected samples (Corbett, 2017). These trained algorithms are fed with thousands of previous case studies and images of medical tests to output unbiased, quick suggestions for doctors, and support their diagnoses.
Hospital-Acquired Infections (HAIs) have always plagued hospitals simply because these institutes are an accumulation of ill patients. HAIs are defined as nosocomial acquired infections that usually present symptoms 48 hours after admission, with conditions ranging from central line-associated bloodstream infections, surgical site infections, Pneumonia, to other common bacterial infections (Monegro, Muppidi and Regunath, 2020). It has been estimated that four percent of all admitted patients experience HAIs, which would lead to readmission, an extended hospital stay, or increased complications. The CDC has estimated a cost of 97 to 147 billion US dollars attributed to direct and indirect cost of HAIs annually, a potentially reducible sinkhole (Healthcare associated infections (HAIs), 2019). Early efforts against HAI used tracking with limited success, halted by the lack of computational power and simplicity of the implementation.
It is only fair to foresee the potential of Machine Learning in these applications. The abundance of patient information collected provides a strong foundation for robust machine learning algorithms. To imagine the large amount of data required for such an endeavor, patient electronic records provide basic information ranging from gender, age, and preexisting conditions and in addition to patient-specific data like symptoms, course of treatment, given vaccinations, and up to 50 other potential factors (Luz, Vollmer and Sinha, 2020). This would constitute a single dataset among the necessary tens of thousands required for prediction algorithms. With these programs dealing with the well-being of humans, high accuracy thresholds might further require hundreds of thousands of additional datasets, ensuring maximum reliability.
Area Under the Receiver Operating Characteristics curve (AUROC) is the accuracy measurement threshold that is the conventional benchmark for these algorithms. These areas under the curve values are calculated graphing the false-positive rates against true positive rates, which indicate specificity and sensitivity of trained algorithms, where values nearer to one indicate a robust algorithm. For HAIs, certain algorithms have reached an impressive 0.92 AUROC score. For other targeted outcomes like sepsis, postoperative infections, and microbiological infections accuracy rates of trained algorithms have reached 0.96, 0.96, and 0.77 respectively (Nasman, 2018). It is crucial to note that different accuracy thresholds are required depending on the targeted outcome and algorithms utilize different machine learning models to achieve them. The exceeding complex field of machine learning offers a plethora of models ranging from artificial neural networks, support vector machines, extreme gradient boosting, to regularized logistic regression, which this crude list hardly does justice to the beauty of its sophistication. Each model has its specific set of assumptions and methodology to process the input data, patient datasets, into an output prediction. Fortunately, this enormous task has seen promising results with future attempts being boosted parallel with the exponential growth of technology.
The path to practically incorporating these machine learning algorithms into the healthcare system might seem straightforward, but is shrouded by the fog of skepticism. As with all new introductions to the healthcare system, from vaccinations to personalized medicine to uses of medical cannabinoids, trust within the industry is earned not given. If one hundred percent accuracy is not realistic, how high of an AUROC would be considered safe to apply to human lives? Can computer programs provide the same, or better, quality diagnosis as a medical professional with years of experience? The winding journey ahead to attaining a robust, reliable algorithm is followed by the grueling test of the healthcare system’s reluctant view.
Scaling the mountain that is the healthcare industry requires incremental incorporations of machine learning in low-risk situations that will build patient trust. Reminding patients to take their medicine, prioritizing patients in order of severity, and notifying of harmful cross drug interactions are some of the avoidable risks simple machine programs have reduced (Srikanth, 2019). Imagine the far-reaching potential of machine learning if it could tackle the bigger medical monsters that have plagued humanity.
DocPanel. 2020. How Common Is Misdiagnosis – Infographic. [online] Available at: <https://www.docpanel.com/blog/post/how-common-misdiagnosis-infographic#:~:text=Each%20year%20in%20the%20U.S.,out%20of%2020%20adult%20patients> [Accessed 1 November 2020].
Corbett, E., 2017. Real-World Benefits Of Machine Learning In Healthcare. [online] Health Catalyst. Available at: <https://www.healthcatalyst.com/clinical-applications-of-machine-learning-in-healthcare> [Accessed 1 November 2020].
Monegro, A., Muppidi, V. and Regunath, H., 2020. Hospital Acquired Infections. [online] Ncbi.nlm.nih.gov. Available at: <https://www.ncbi.nlm.nih.gov/books/NBK441857/> [Accessed 1 November 2020].
Premiersafetyinstitute.org. 2019. Healthcare Associated Infections (Hais) |. [online] Available at: <https://www.premiersafetyinstitute.org/safety-topics-az/healthcare-associated-infections-hais/hai/> [Accessed 1 November 2020].
Luz, C., Vollmer, M. and Sinha, B., 2020. Machine Learning In Infection Management Using Routine Electronic Health Records: Tools, Techniques, And Reporting Of Future Technologies. https://www.sciencedirect.com/science/article/pii/S1198743X20300823.
NÄSMAN, M., 2018. Detecting Hospital Acquired Infections Using Machine Learning. [online] Diva-portal.org. Available at: <http://www.diva-portal.org/smash/get/diva2:699767/FULLTEXT01.pdf> [Accessed 1 November 2020].
Srikanth, J., 2019. 15 Benefits Of Machine Learning In Health Care – Techiexpert.Com. [online] Techiexpert.com. Available at: <https://www.techiexpert.com/benefits-of-machine-learning-in-health-care/> [Accessed 1 November 2020].