Written by: Abhay Mahajan
Edited by: Ryan Lee
Edited by: Ryan Lee
As our world becomes increasingly digitized, even the field of medicine is turning towards machine learning. Machine learning (ML) involves supplying an artificial intelligence system with data in order to help the system recognize certain patterns. By feeding these systems large amounts of data, they can be used to describe and predict certain patterns of interest as well as offer potential solutions or actions in response to the observations (Brown 2021). It is these features which have captured the interests of medical researchers, with artificial intelligence (AI) perhaps being the next breakthrough in improving the accessibility of medical care as well as increasing the speed at which we can diagnose different ailments.
In the case of cardiovascular disease, this technology comes with great implications. Currently, heart disease is the leading cause of death for both men and women within the U.S., with roughly 697 thousand dying of heart disease within 2020 alone (CDC 2022). With the use of wearable sensors to continually monitor heart rate, heart rhythm, blood oxygen saturation, and other data to diagnose atrial fibrillation (AF) and other cardiac events, AI can serve as a strong ally in combating the lethality of heart disease (Huang et al. 2022).
In fact, AI was suggested to be able to predict AF events four hours beforehand through the use of photoplethysmography (PPG) data (Huang et al. 2022). PPG tests evaluate the blood waveforms and blood pressure through the use of a blood pressure probe around the toe, finger, etc. (Stanford 2022). Yutao Guo et al. performed PPG’s on Chinese patients 18 years or older via monitoring wristbands and wristwatches. Using PPG data, the algorithm would determine if AF was suspected. In the case where AF was not suspected, an analysis of the measurement frequency and irregular pulse rhythm would be conducted daily.. If AF was not detected over two weeks, the individual would be flagged as not being suspected of having AF. The findings of the algorithm were further confirmed by health care providers. In addition to this, the researchers implemented a discrimination rule with the variable threshold “T” to determine if a rate of at least “T” AF events occurred over the course of more than 10 PPG measurements; if this was the case, AF would be suspected. This allowed for a positive predictive value of over 85% (Guo et al. 2019).
The data from this study was then used by Guo et al. to create a neural network based on supervised learning. Supervised learning is a technique in which a ML program is made to output certain data based off of given inputs . This is in contrast to unsupervised learning where not every input may result in an output. This is typically used to identify unknown patterns with the given data. For example, a video service may gather data on which videos you watch and eventually discover a pattern in the types of content you may enjoy. The diagnosis of AF is possible because AF correlates with atrial fibrillation burden, or the proportion of monitoring time that one is experiencing AF (Huang et al. 2022). Additionally, the ability for AI to predict AF is due to AF correlating with 17 different features of PPG signals, such as heart rate and heart rate variability (Guo et al. 2021). By monitoring a patient’s heart through the use of wearable EKG patches or PPG devices, it is possible that physicians may be able to get the patient proper medical attention earlier, thus lowering the potential complications of AF. It is especially important to diagnose AF in a timely fashion as a missed diagnosis may increase the risk of stroke (NINDS 2022).
Similarly, machine learning can be applied in the case of a myocardial infarction, or heart attack. Using a Bidirectional Long Short-Term Memory (BiLSTM) model, a model used for neural networks to make predictions based on given data, the software was able to detect ST-elevation myocardial infarction (STEMI) with an accuracy of 98.7% (Huang et al. 2022). This is especially important because STEMIs are one of the more aggressive forms of myocardial infarctions and have a high mortality and morbidity rate if left untreated (Cleveland Clinic 2021). Despite this high level of accuracy for detecting STEMIs, it is still unclear whether STEMIs can be predicted beforehand. This is because STEMIs are caused due to the complete blockage of a coronary artery, resulting in the death of cardiac cells (Cleveland Clinic 2021). This blockage is hard to predict as it would require the monitoring of coronary artery plaque which may not even be present upon initial measurement. However, researchers at Cedars Sinai are aiming to create a software capable of predicting myocardial infarction (Cedars Sinai 2022).
Artificial intelligence has also proven itself useful for detecting a variety of oral diseases. The AI uses convolutional neural networks (CNN), which are useful for detecting visual patterns, to detect diseases such as dental caries and tooth fracture with great accuracy. By supplying the network with 3686 radiographs, with 3293 being used for training and 252 for test data, the algorithm was able to detect dental caries more accurately than dentists (Patil et al. 2022). This ML model works by feeding the CNN near-infrared transillumination images and labeling the initial samples as having dental caries or not, with the final set of data being used to test the predictive accuracy of the neural network (Patil et al. 2022). In the case of tooth fractures, machine learning models were capable of detecting fractures with high specificity and sensitivity (Patil et al. 2022).
This demonstrates that AI systems show great potential for helping medical professionals detect disease. However, there are still many limitations that exist in terms of collecting the necessary data. For example, in the case of rare diseases, it may be hard to obtain samples large enough for an AI network to be able to accurately recognize patterns associated with the disease. Additionally, there may not yet be determined characteristics which correlate with certain diseases or these characteristics may only show themselves for a short period of time relative to the onset of the disease, such as in the case of myocardial infarction. Nonetheless, AI is likely to become a major factor in healthcare of the future, with machine learning already showing great potential for predicting and detecting a variety of diseases.
References
MIT Management [Internet]. 2021. Massachusetts, U.S.: MIT; [updated Apr 21, 2021; cited Nov 17, 2022]. Available from: https://mitsloan.mit.edu/ideas-made-to-matter/machine-learning-explained#:~:text=What%20is%20machine%20learning%3F,to%20how%20humans%20solve%20problems
Centers for Disease Control and Prevention [Internet]. CDC;[updated Oct 14, 2021; cited December 8, 2022]. Available from: https://www.cdc.gov/heartdisease/facts.htm
Cedars Sinai [Internet]. 2022. Los Angeles, CA: Cedars Sinai;[updated Mar 22, 2022; cited Dec 8, 2022]. Available from: https://www.cedars-sinai.org/newsroom/artificial-intelligence-tool-may-help-predict-heart-attacks/
Cleveland Clinic [Internet]. Cleveland Clinic; [updated Nov 15, 2021; cited Dec 6, 2022]. Available from: https://my.clevelandclinic.org/health/diseases/22068-stemi-heart-attack
Guo Y, Wang H, Zhang H, Lou T, Liang Z, Xia Y, Yan L, Xing Y, Shi H, Li S, et al. 2019. Mobile Photoplethysmographic Technology to Detect Atrial Fibrillation. Journal of the American College of Cardiology [Internet]. [updated Nov 12, 2019; cited Nov 17, 2022]; 74(19): 2365-2375. Available from: https://doi.org/10.1016/j.jacc.2019.08.019
Guo Y, Wang H, Zhang H, Lou T, Li L, Lou L, Chen M, Chen Y, Lip GYH. 2021. Photoplethysmography-Based Machine Learning Approaches for Atrial Fibrillation Prediction: A Report From the Huawei Heart Study. Journal of the American College of Cardiology: Asia [Internet]. [cited Dec 6, 2022]; 1(3): 399-408. Available from: https://doi.org/10.1016/j.jacasi.2021.09.004
Huang JD, Wang J, Ramsey E, Leavey G, Chico T, Condell J. Applying Artificial Intelligence to Wearable Sensor Data to Diagnose and Predict Cardiovascular Disease: A Review. Sensors [Internet]. [cited Nov 17, 2022]; 22(20): 28. Available from DOI: 10.3390/s22208002
National Institute of Neurological Disorders and Stroke [Internet]. National Institute of Health; [updated Jul 25, 2022, cited Dec 8, 2022]. Available from: https://www.ninds.nih.gov/health-information/disorders/atrial-fibrillation-and-stroke
Patil S, Albogami S, Hosmani J, Mujoo S, Kami MA, Mansour MA, Abdul HN, Bhandi S, Ahmed S. Artificial Intelligence in the Diagnosis of Oral Diseases: Applications and Pitfalls. Diagnostics [Internet]. [cited Nov 28, 2022]; 12(5): 14. Available from DOI: 10.3390/diagnostics12051029
Stanford Medicine [Internet]. [cited Nov 27, 2022]. Available from: https://stanfordhealthcare.org/medical-tests/p/ppg.html
In the case of cardiovascular disease, this technology comes with great implications. Currently, heart disease is the leading cause of death for both men and women within the U.S., with roughly 697 thousand dying of heart disease within 2020 alone (CDC 2022). With the use of wearable sensors to continually monitor heart rate, heart rhythm, blood oxygen saturation, and other data to diagnose atrial fibrillation (AF) and other cardiac events, AI can serve as a strong ally in combating the lethality of heart disease (Huang et al. 2022).
In fact, AI was suggested to be able to predict AF events four hours beforehand through the use of photoplethysmography (PPG) data (Huang et al. 2022). PPG tests evaluate the blood waveforms and blood pressure through the use of a blood pressure probe around the toe, finger, etc. (Stanford 2022). Yutao Guo et al. performed PPG’s on Chinese patients 18 years or older via monitoring wristbands and wristwatches. Using PPG data, the algorithm would determine if AF was suspected. In the case where AF was not suspected, an analysis of the measurement frequency and irregular pulse rhythm would be conducted daily.. If AF was not detected over two weeks, the individual would be flagged as not being suspected of having AF. The findings of the algorithm were further confirmed by health care providers. In addition to this, the researchers implemented a discrimination rule with the variable threshold “T” to determine if a rate of at least “T” AF events occurred over the course of more than 10 PPG measurements; if this was the case, AF would be suspected. This allowed for a positive predictive value of over 85% (Guo et al. 2019).
The data from this study was then used by Guo et al. to create a neural network based on supervised learning. Supervised learning is a technique in which a ML program is made to output certain data based off of given inputs . This is in contrast to unsupervised learning where not every input may result in an output. This is typically used to identify unknown patterns with the given data. For example, a video service may gather data on which videos you watch and eventually discover a pattern in the types of content you may enjoy. The diagnosis of AF is possible because AF correlates with atrial fibrillation burden, or the proportion of monitoring time that one is experiencing AF (Huang et al. 2022). Additionally, the ability for AI to predict AF is due to AF correlating with 17 different features of PPG signals, such as heart rate and heart rate variability (Guo et al. 2021). By monitoring a patient’s heart through the use of wearable EKG patches or PPG devices, it is possible that physicians may be able to get the patient proper medical attention earlier, thus lowering the potential complications of AF. It is especially important to diagnose AF in a timely fashion as a missed diagnosis may increase the risk of stroke (NINDS 2022).
Similarly, machine learning can be applied in the case of a myocardial infarction, or heart attack. Using a Bidirectional Long Short-Term Memory (BiLSTM) model, a model used for neural networks to make predictions based on given data, the software was able to detect ST-elevation myocardial infarction (STEMI) with an accuracy of 98.7% (Huang et al. 2022). This is especially important because STEMIs are one of the more aggressive forms of myocardial infarctions and have a high mortality and morbidity rate if left untreated (Cleveland Clinic 2021). Despite this high level of accuracy for detecting STEMIs, it is still unclear whether STEMIs can be predicted beforehand. This is because STEMIs are caused due to the complete blockage of a coronary artery, resulting in the death of cardiac cells (Cleveland Clinic 2021). This blockage is hard to predict as it would require the monitoring of coronary artery plaque which may not even be present upon initial measurement. However, researchers at Cedars Sinai are aiming to create a software capable of predicting myocardial infarction (Cedars Sinai 2022).
Artificial intelligence has also proven itself useful for detecting a variety of oral diseases. The AI uses convolutional neural networks (CNN), which are useful for detecting visual patterns, to detect diseases such as dental caries and tooth fracture with great accuracy. By supplying the network with 3686 radiographs, with 3293 being used for training and 252 for test data, the algorithm was able to detect dental caries more accurately than dentists (Patil et al. 2022). This ML model works by feeding the CNN near-infrared transillumination images and labeling the initial samples as having dental caries or not, with the final set of data being used to test the predictive accuracy of the neural network (Patil et al. 2022). In the case of tooth fractures, machine learning models were capable of detecting fractures with high specificity and sensitivity (Patil et al. 2022).
This demonstrates that AI systems show great potential for helping medical professionals detect disease. However, there are still many limitations that exist in terms of collecting the necessary data. For example, in the case of rare diseases, it may be hard to obtain samples large enough for an AI network to be able to accurately recognize patterns associated with the disease. Additionally, there may not yet be determined characteristics which correlate with certain diseases or these characteristics may only show themselves for a short period of time relative to the onset of the disease, such as in the case of myocardial infarction. Nonetheless, AI is likely to become a major factor in healthcare of the future, with machine learning already showing great potential for predicting and detecting a variety of diseases.
References
MIT Management [Internet]. 2021. Massachusetts, U.S.: MIT; [updated Apr 21, 2021; cited Nov 17, 2022]. Available from: https://mitsloan.mit.edu/ideas-made-to-matter/machine-learning-explained#:~:text=What%20is%20machine%20learning%3F,to%20how%20humans%20solve%20problems
Centers for Disease Control and Prevention [Internet]. CDC;[updated Oct 14, 2021; cited December 8, 2022]. Available from: https://www.cdc.gov/heartdisease/facts.htm
Cedars Sinai [Internet]. 2022. Los Angeles, CA: Cedars Sinai;[updated Mar 22, 2022; cited Dec 8, 2022]. Available from: https://www.cedars-sinai.org/newsroom/artificial-intelligence-tool-may-help-predict-heart-attacks/
Cleveland Clinic [Internet]. Cleveland Clinic; [updated Nov 15, 2021; cited Dec 6, 2022]. Available from: https://my.clevelandclinic.org/health/diseases/22068-stemi-heart-attack
Guo Y, Wang H, Zhang H, Lou T, Liang Z, Xia Y, Yan L, Xing Y, Shi H, Li S, et al. 2019. Mobile Photoplethysmographic Technology to Detect Atrial Fibrillation. Journal of the American College of Cardiology [Internet]. [updated Nov 12, 2019; cited Nov 17, 2022]; 74(19): 2365-2375. Available from: https://doi.org/10.1016/j.jacc.2019.08.019
Guo Y, Wang H, Zhang H, Lou T, Li L, Lou L, Chen M, Chen Y, Lip GYH. 2021. Photoplethysmography-Based Machine Learning Approaches for Atrial Fibrillation Prediction: A Report From the Huawei Heart Study. Journal of the American College of Cardiology: Asia [Internet]. [cited Dec 6, 2022]; 1(3): 399-408. Available from: https://doi.org/10.1016/j.jacasi.2021.09.004
Huang JD, Wang J, Ramsey E, Leavey G, Chico T, Condell J. Applying Artificial Intelligence to Wearable Sensor Data to Diagnose and Predict Cardiovascular Disease: A Review. Sensors [Internet]. [cited Nov 17, 2022]; 22(20): 28. Available from DOI: 10.3390/s22208002
National Institute of Neurological Disorders and Stroke [Internet]. National Institute of Health; [updated Jul 25, 2022, cited Dec 8, 2022]. Available from: https://www.ninds.nih.gov/health-information/disorders/atrial-fibrillation-and-stroke
Patil S, Albogami S, Hosmani J, Mujoo S, Kami MA, Mansour MA, Abdul HN, Bhandi S, Ahmed S. Artificial Intelligence in the Diagnosis of Oral Diseases: Applications and Pitfalls. Diagnostics [Internet]. [cited Nov 28, 2022]; 12(5): 14. Available from DOI: 10.3390/diagnostics12051029
Stanford Medicine [Internet]. [cited Nov 27, 2022]. Available from: https://stanfordhealthcare.org/medical-tests/p/ppg.html
Proudly powered by Weebly