Machine learning and proteomics predict cardiovascular risk more accurately
DISCLAIMER: This blog is fully automated, unmonitored, and does not reflect the views of Garett MacGowan. The ML model may produce content that is offensive to some readers.
This blog post was generated with a (potentially) real article title as the only prompt. A link to the original article is below.
Photo by Markus Spiske on Unsplash
Generated: 4/11/2022
Machine learning and proteomics predict cardiovascular risk more accurately than standard approaches {#sec0005}
=======================================================================================================
We have entered a new era in medical care, where the power of machine learning (ML) is used to create highly predictive models that provide an unprecedented degree of personalized prediction, diagnosis, and response towards potential adverse disease outcomes. These models are based on the massive amounts of high-dimensional data available and on recent advances in ML algorithms, allowing real-time prediction.
In cardiology, ML models enable us to identify predictors of cardiovascular outcomes and to predict individual disease risk much better than standard approaches, as shown in multiple reports. Here we discuss how the application of ML methods to identify patients at increased risk has major implications for clinical practice as well as for future diagnostic and therapeutic strategies. The development of ML models in clinical cardiology is an exciting new area, and a PubMed search with the terms 'cardiology' and'machine learning' resulted in more than 1400 hits in August 2018 alone, suggesting that this field is expanding rapidly both *in vivo* and *in silico*.
ML in cardiovascular medicine: a brief historical perspective {#sec0010}
=============================================================
Cardiovascular disease (CVD) is the number one cause of death in developed and developing countries, surpassing even cancer [@bib0135]. Since the first studies investigating ML in clinical cardiology, we have seen an explosion of ML models designed to provide personalized predictions to improve diagnosis, management of patients, as well as therapeutic decision making and to identify new therapeutic targets and biomarkers.
Early work with ML applied artificial neural networks to predict myocardial infarction from patient variables, such as blood markers, Holter recordings, and symptoms. Most of the studies reported a specificity of the model better than 95% with a sensitivity between 40 and 60% [@bib0140], [@bib0145], [@bib0150], [@bib0155], [@bib0160], [@bib0165]. In the clinical setting, the best models to predict myocardial infarction were applied to a large heterogeneous group of patients without personalized treatment, and a large number of patients (\>75,000) would be required to achieve the same degree of accuracy offered by ML models used in the laboratory. Additionally, several authors were able to show that ML models could predict the development of heart failure over a 3-year period from standard baseline variables with accuracy levels close to 80% [@bib0140], [@bib0165].
In the last decade, big data and novel, powerful ML algorithms have revolutionized ML methods. In 2016, one of the first studies to use big data and ML was published by Tzikas et al. [@bib0150], demonstrating that patients with heart failure had a high score of the Framingham 10-year risk of a cardiovascular event (F-10), a risk tool commonly used by cardiologists to predict cardiovascular events. The F-10 score was based on age, sex, cholesterol levels, blood pressure, diabetes, smoking status, left ventricle ejection fraction, pulse wave velocity, body mass index, and the presence of heart failure, all of which were taken from routinely recorded patient data. Interestingly, at this early point in cardiology, heart failure developed on average 16 years before the end-organ damage of the disease could be detected based on ejection fraction loss, thereby demonstrating the advantage of ML algorithms using big data.
In subsequent work, similar observations of excellent prognostic performance of simple F-10 and F-10 plus traditional CVD risk factors using ML models were confirmed in several reports investigating various cardiovascular diseases, such as myocardial infarction [@bib0145], [@bib0155], [@bib0160], hypertension [@bib0160], heart failure [@bib0150], [@bib0170], and arterial atherosclerosis [@bib0140]. These were some of the first examples of ML applied to prognostic prediction, highlighting its potential to improve clinical care of patients and, as such, has the potential to have a major impact on CVD and cardiovascular medicine.
ML in cardiovascular medicine: more recent developments using big data {#sec0015}
======================================================================
Over the last few years, an explosion of studies using big data to enable early diagnosis has been reported. In 2014, an innovative ML approach was used to predict the risk for hospitalization for myocardial infarction in patients presenting to an Emergency Department (ED) [@bib0175]. Data from more than 15,000 presentations to the ED of a single city in the UK were analyzed. The study included several variables previously shown to be associated with the risk of adverse event development, such as elevated levels of high-sensitivity cardiac troponin I [@bib0175]. ML was able to identify previously unanticipated and independent relationships between variables and events such as age, sex, and elevated troponin or creatinine levels, demonstrating its potential to reduce patient exposure to ionizing radiation for cardiac testing and its potential applicability to decision making for emergent or urgent care.
Next, Wang et al. [@bib0155] used a combination of traditional CVD risk factors and ML in 10,000 patients without recent myocardial infarction to predict major adverse cardiac events in patients with a previously confirmed ischemic stroke or transient ischemic attack, showing an improved risk score accuracy compared to F-10 and F-10 plus other variables [@bib0155].
Although these early studies focused on risk prediction using ML models, a few groups have investigated disease classification. Bhattacharyya et al. [@bib0180] conducted a large multicentre study to classify patients with suspected myocardial infarction into two groups, unstable angina and myocardial infarction. Although the accuracy for the diagnosis was only 74% for ML models, important associations were identified with novel biological markers such as the soluble form of the urokinase-type plasminogen activator receptor, which demonstrated prognostic values in other studies [@bib0185], [@bib0190]. Another example of ML model development from data with a previously described outcome was the recent report by Cai et al. [@bib0195].
The aforementioned studies all used big data and ML to build models. A more recent study investigated the prognostic potential of previously published ML models from the Framingham Heart Study that had been built using very simple variables such as age, sex, cholesterol, and blood pressure. The models identified the 30-year probability of a cardiovascular endpoint with an accuracy \>66% [@bib0200].
A recent study explored the predictive power of ML in real patients treated at a specialized heart failure outpatient clinic in a hospital-based prospective study [@bib0065]. The authors used variables available from routine patient care and assessed the prognostic ability of ML methods based on various ML algorithms such as a regression tree and logistic regression as well as ML models developed based on big data with ML using a cohort of 20,000 patients [@bib0065]. In this study, the ML models demonstrated superior predictability of mortality at 12 months of follow-up compared with traditional algorithms based on the same dataset [@bib0065].
Interestingly, the models identified very simple variables as independent predictors of a poor outcome. These were hemogloblin concentration \<12 g/dL and higher serum albumin concentrations (\>36.9 g/dL), which are likely markers of the severity of heart failure or a poor general state of health. These variables were found to have prognostic power comparable to the best ML models [@bib0065] ([Fig. 2](#fig0010){ref-type="fig"}).Fig. 2ML approaches as an aid to prognosis or diagnosis. (A) The majority of studies use machine learning (ML) to identify predictors of adverse cardiovascular events or to predict cardiac events according to a large database of a large number of patients. The models are then used to calculate and predict event rates using a regression model. (B) Several studies use ML to distinguish between patients with different pathophysiologic disease states. This information may be used to determine whether certain patients are candidates for certain therapeutic approaches or are likely not to respond to a therapy. A similar approach may be used to identify patients likely to develop cardiac events.Fig. 2
The majority of studies that have used ML to identify predictors of adverse events in patients are based on patient data, which is typically collected by hospital staff (e.g., doctors and nurses) or by family members. Nevertheless, it is unlikely that ML will replace doctors, as it only assists doctors to improve their diagnostic and prognostic abilities and does not aim to replace human medical knowledge, nor can the ML tools predict new diseases (personalized medicine). Thus, ML technology will only help clinicians detect, diagnose, or suggest therapies for diseases that currently would go undiagnosed or undiscussed, and could even reduce costs [@bib0205]. Most importantly, a high-level of ML requires a huge amount of data: the more data that are available, the more accurate the results obtained from the ML.
ML models versus current approaches {#sec0020}
===================================
Even in current cardiology, several studies have suggested that ML approaches could be more accurate than commonly used methods. In this section, we review some of these studies, some in which ML predicted disease outcomes ([Fig. 2A](#fig0010){ref-type="fig"}) and others in which they discriminated patients into different states of illness ([Fig.
=======================================================================================================
We have entered a new era in medical care, where the power of machine learning (ML) is used to create highly predictive models that provide an unprecedented degree of personalized prediction, diagnosis, and response towards potential adverse disease outcomes. These models are based on the massive amounts of high-dimensional data available and on recent advances in ML algorithms, allowing real-time prediction.
In cardiology, ML models enable us to identify predictors of cardiovascular outcomes and to predict individual disease risk much better than standard approaches, as shown in multiple reports. Here we discuss how the application of ML methods to identify patients at increased risk has major implications for clinical practice as well as for future diagnostic and therapeutic strategies. The development of ML models in clinical cardiology is an exciting new area, and a PubMed search with the terms 'cardiology' and'machine learning' resulted in more than 1400 hits in August 2018 alone, suggesting that this field is expanding rapidly both *in vivo* and *in silico*.
ML in cardiovascular medicine: a brief historical perspective {#sec0010}
=============================================================
Cardiovascular disease (CVD) is the number one cause of death in developed and developing countries, surpassing even cancer [@bib0135]. Since the first studies investigating ML in clinical cardiology, we have seen an explosion of ML models designed to provide personalized predictions to improve diagnosis, management of patients, as well as therapeutic decision making and to identify new therapeutic targets and biomarkers.
Early work with ML applied artificial neural networks to predict myocardial infarction from patient variables, such as blood markers, Holter recordings, and symptoms. Most of the studies reported a specificity of the model better than 95% with a sensitivity between 40 and 60% [@bib0140], [@bib0145], [@bib0150], [@bib0155], [@bib0160], [@bib0165]. In the clinical setting, the best models to predict myocardial infarction were applied to a large heterogeneous group of patients without personalized treatment, and a large number of patients (\>75,000) would be required to achieve the same degree of accuracy offered by ML models used in the laboratory. Additionally, several authors were able to show that ML models could predict the development of heart failure over a 3-year period from standard baseline variables with accuracy levels close to 80% [@bib0140], [@bib0165].
In the last decade, big data and novel, powerful ML algorithms have revolutionized ML methods. In 2016, one of the first studies to use big data and ML was published by Tzikas et al. [@bib0150], demonstrating that patients with heart failure had a high score of the Framingham 10-year risk of a cardiovascular event (F-10), a risk tool commonly used by cardiologists to predict cardiovascular events. The F-10 score was based on age, sex, cholesterol levels, blood pressure, diabetes, smoking status, left ventricle ejection fraction, pulse wave velocity, body mass index, and the presence of heart failure, all of which were taken from routinely recorded patient data. Interestingly, at this early point in cardiology, heart failure developed on average 16 years before the end-organ damage of the disease could be detected based on ejection fraction loss, thereby demonstrating the advantage of ML algorithms using big data.
In subsequent work, similar observations of excellent prognostic performance of simple F-10 and F-10 plus traditional CVD risk factors using ML models were confirmed in several reports investigating various cardiovascular diseases, such as myocardial infarction [@bib0145], [@bib0155], [@bib0160], hypertension [@bib0160], heart failure [@bib0150], [@bib0170], and arterial atherosclerosis [@bib0140]. These were some of the first examples of ML applied to prognostic prediction, highlighting its potential to improve clinical care of patients and, as such, has the potential to have a major impact on CVD and cardiovascular medicine.
ML in cardiovascular medicine: more recent developments using big data {#sec0015}
======================================================================
Over the last few years, an explosion of studies using big data to enable early diagnosis has been reported. In 2014, an innovative ML approach was used to predict the risk for hospitalization for myocardial infarction in patients presenting to an Emergency Department (ED) [@bib0175]. Data from more than 15,000 presentations to the ED of a single city in the UK were analyzed. The study included several variables previously shown to be associated with the risk of adverse event development, such as elevated levels of high-sensitivity cardiac troponin I [@bib0175]. ML was able to identify previously unanticipated and independent relationships between variables and events such as age, sex, and elevated troponin or creatinine levels, demonstrating its potential to reduce patient exposure to ionizing radiation for cardiac testing and its potential applicability to decision making for emergent or urgent care.
Next, Wang et al. [@bib0155] used a combination of traditional CVD risk factors and ML in 10,000 patients without recent myocardial infarction to predict major adverse cardiac events in patients with a previously confirmed ischemic stroke or transient ischemic attack, showing an improved risk score accuracy compared to F-10 and F-10 plus other variables [@bib0155].
Although these early studies focused on risk prediction using ML models, a few groups have investigated disease classification. Bhattacharyya et al. [@bib0180] conducted a large multicentre study to classify patients with suspected myocardial infarction into two groups, unstable angina and myocardial infarction. Although the accuracy for the diagnosis was only 74% for ML models, important associations were identified with novel biological markers such as the soluble form of the urokinase-type plasminogen activator receptor, which demonstrated prognostic values in other studies [@bib0185], [@bib0190]. Another example of ML model development from data with a previously described outcome was the recent report by Cai et al. [@bib0195].
The aforementioned studies all used big data and ML to build models. A more recent study investigated the prognostic potential of previously published ML models from the Framingham Heart Study that had been built using very simple variables such as age, sex, cholesterol, and blood pressure. The models identified the 30-year probability of a cardiovascular endpoint with an accuracy \>66% [@bib0200].
A recent study explored the predictive power of ML in real patients treated at a specialized heart failure outpatient clinic in a hospital-based prospective study [@bib0065]. The authors used variables available from routine patient care and assessed the prognostic ability of ML methods based on various ML algorithms such as a regression tree and logistic regression as well as ML models developed based on big data with ML using a cohort of 20,000 patients [@bib0065]. In this study, the ML models demonstrated superior predictability of mortality at 12 months of follow-up compared with traditional algorithms based on the same dataset [@bib0065].
Interestingly, the models identified very simple variables as independent predictors of a poor outcome. These were hemogloblin concentration \<12 g/dL and higher serum albumin concentrations (\>36.9 g/dL), which are likely markers of the severity of heart failure or a poor general state of health. These variables were found to have prognostic power comparable to the best ML models [@bib0065] ([Fig. 2](#fig0010){ref-type="fig"}).Fig. 2ML approaches as an aid to prognosis or diagnosis. (A) The majority of studies use machine learning (ML) to identify predictors of adverse cardiovascular events or to predict cardiac events according to a large database of a large number of patients. The models are then used to calculate and predict event rates using a regression model. (B) Several studies use ML to distinguish between patients with different pathophysiologic disease states. This information may be used to determine whether certain patients are candidates for certain therapeutic approaches or are likely not to respond to a therapy. A similar approach may be used to identify patients likely to develop cardiac events.Fig. 2
The majority of studies that have used ML to identify predictors of adverse events in patients are based on patient data, which is typically collected by hospital staff (e.g., doctors and nurses) or by family members. Nevertheless, it is unlikely that ML will replace doctors, as it only assists doctors to improve their diagnostic and prognostic abilities and does not aim to replace human medical knowledge, nor can the ML tools predict new diseases (personalized medicine). Thus, ML technology will only help clinicians detect, diagnose, or suggest therapies for diseases that currently would go undiagnosed or undiscussed, and could even reduce costs [@bib0205]. Most importantly, a high-level of ML requires a huge amount of data: the more data that are available, the more accurate the results obtained from the ML.
ML models versus current approaches {#sec0020}
===================================
Even in current cardiology, several studies have suggested that ML approaches could be more accurate than commonly used methods. In this section, we review some of these studies, some in which ML predicted disease outcomes ([Fig. 2A](#fig0010){ref-type="fig"}) and others in which they discriminated patients into different states of illness ([Fig.