As we know, multiple medications exist for the same condition — however, how do we know which drug is suited to which individual? Machine learning is evolving and advancing in its utility, with various studies exploring how it might aid the personalised assessment and treatment of individual patients.
Sotomi et al (2023) examine the use of machine learning to identify cohorts of a population that may benefit from specific medications. The team had previously established a machine learning-based clustering model that was able to classify heart failure with preserved ejection fraction (HFpEF) into four distinct phenotypes.
Individuals may respond differently to medications based on their phenotype. Therefore, specific drugs may be more effective in treating each of the four phenotypes. To be able to determine if this is the case, the researchers conducted an assessment. If successful, this approach could have a positive impact on the treatment of certain populations with heart failure. By identifying the appropriate medication at an earlier stage, long-term outcomes could be improved.
Sotomi et al (2023) conducted a post-hoc analysis of the PURSUIT-HFpEF registry, which was a prospective, multicenter, observational study. The researchers evaluated the clinical effectiveness of four different types of medication given post-discharge, across the four different phenotypes. These medications were angiotensin-converting enzyme inhibitors (ACEi) or angiotensin-receptor blockers (ARB), beta blockers, mineralocorticoid-receptor antagonists (MRA) and statins.
The research team analysed the primary endpoint of all-cause death and hospitalisation as a result of heart failure in a sample of 1231 patients, with a median follow-up duration of 734 days. Of the sample, 48% (528 patients) experienced the primary endpoint. The team determined that learning-based clustering has the potential to identify populations for which specific medications may be effective. For example, the study found evidence that MRA, ACEi or ARB, and statins were effective in treating specific phenotypes of HFpEF.
Sotomi et al (2023) summarised their findings regarding the clinical effectiveness of four different medications across the four phenotypes of heart failure with preserved ejection fraction (HFpEF). The researchers noted that machine learning-based clustering has the potential to identify populations in which specific medications may be effective. They also found that none of the four medications evaluated in the analysis had a significant effect on clinical outcomes in phenotype 1. In contrast, MRA therapy significantly improved clinical outcomes in phenotype 2. ACEi or ARB and statin therapy significantly improved clinical outcomes in phenotype 3. Finally, beta blockers tended to worsen the clinical outcomes in phenotype 4. These findings suggest that personalised treatment strategies based on phenotype may improve clinical outcomes for patients with HFpEF.
Sotomi et al (2023) described phenotype 1 as ‘rhythm trouble’. This phenotype is characterised by a low level of comorbidity and primarily worsens heart failure through atrial fibrillation, which typically has a benign prognosis. According to the research team, ACEi/ARB blockers, beta blockers, MRA and statins were mostly lacking in efficacy for this group of patients. Instead, they suggested that patients with phenotype 1 may only benefit from a combination of antiarrhythmic drugs and catheter ablation to manage their condition.
Sotomi et al (2023) found that MRA was particularly significant in improving outcomes for patients with phenotype 2, which is characterised by cardiac hypertrophy and hypertension, common features in HFpEF, and referred to as ‘ventricular-arterial uncoupling’. This phenotype is often the focus of research, with inflammation being the primary element of its pathophysiological development. The authors noted that their previous research had found eplerenone to be effective in reducing structural alteration and diastolic dysfunction, independent of lower blood pressure, which may be why MRA, but not ACEi/ARB, significantly improved clinical outcomes in this phenotype.
The team also suggested that MRA may be effective in blocking upregulated myocardial mineralocorticoid receptors in this particular phenotype. They cited other research that found MRA to be exclusively effective in the phenogroup characterised by obesity, diabetes, chronic kidney disease (CKD), concentric left-ventricular (LV) hypertrophy, high renin, biomarkers of tumour necrosis factor-alpha-mediated inflammation, liver fibrosis and tissue remodeling.
Phenotype 3, described by Sotomi et al (2023) as ‘low output and systemic congestion’, benefited from ACEi/ARB and statin therapy. The group was mainly characterised by people with high rates of CKD and frailty. ACEi or ARB are very well studied medications for hypertension, showing good efficacy for renal and cardiovascular protection in chronic kidney disease. Frail patients were previously seen to benefit from ACEi/ARB, but it is difficult to identify the exact mechanism behind its efficacy in frailty and CKD, which is likely multifactorial. Statin therapy in HFpEF has never been evaluated in a randomised trial, but there have been some observational studies describing its effectiveness in this group of patients. The effectiveness of statins in phenotype 3 was unexpected, as this group carries a relatively low burden of comorbidities (hypertension, diabetes and dyslipidaemia).
The good efficacy of statin therapy may be because of its improvement of endothelial function, an increase in arterial distensibility, regression of cardiac hypertrophy and fibrosis, and anti-inflammatory and immunomodulatory effects. However, Sotomi et al (2023) explain that these mechanisms do not fully explain the specific effectiveness of these drugs in phenotype 3. The team notes that the specific effect of statins on pulmonary hypertension may partially explain why there is significant effectiveness in phenotype 3, as this phenotype had the highest level of pulmonary artery systolic pressure.
Phenotype 4, as described by Sotomi et al (2023), was found to be unresponsive to beta blockers and in fact, this medication may worsen the clinical outcomes of this particular group of patients. The phenotype is characterised by poor nutritional status, high levels of frailty and an increased risk of infection-triggered heart failure. This increased risk may be because of the reflected pressure waves that cause an increase in central blood pressure. Prolonged diastolic filling causes ventricular volumes and pressures to increase, which increases ventricular load and leads to elevated levels of BNP and NT-proBNP. The authors noted that the reason why beta blockers worsened outcomes exclusively in phenotype 4 remains unclear. However, they speculate that the patients in this phenotype may be more prone to chronotropic incompetence because of a higher proportion of elderly with a higher frailty score and worse nutritional status. The use of beta blockers may further decrease cardiac sympathetic activity, which could worsen their chronotropic incompetence and may be the reason why the use of beta blockers in this group resulted in a worse prognosis.
It is important to note that these findings should be interpreted with caution, as many of them are based on hypotheses and only four types of drugs were evaluated. Additionally, the study did not assess the effects of cardiac rehabilitation and self-management. Furthermore, the mechanisms underlying the effectiveness of the identified treatments need to be further explored. Nevertheless, the study suggests that machine learning-based clustering could be a useful tool for identifying patient populations that may benefit from specific medications.
Treating major depressive disorder
A recent study published in Translational Psychiatry explored the use of machine learning to optimise treatment for major depressive disorder (MDD). According to the authors, MDD is a complex condition that poses a significant challenge in determining optimal medication for each patient. Treatment often relies on a trial-and-error approach, with only around half of patients showing a good response to medication.
The study by Taliaz et al (2021) aimed to develop an accurate predictor of response to a panel of antidepressants and optimise treatment selection using a data-driven approach that analyses combinations of genetic, clinical and demographic factors. The team analysed patient response patterns to three different antidepressant medications in the Sequenced Treatment Alternatives to Relieve Depression (STAR*D) study, using state-of-the-art machine learning tools to generate a predictive algorithm.
To validate their results, the team assessed the algorithm's capacity to predict individualised antidepressant responses on a separate set of 530 patients in STAR*D, which included 271 patients in a validation set and 259 patients in the final test set. Their assessment produced an average balanced accuracy rate of 72.3% and 70.1% across the different medications in the validation and test set, respectively. To further validate the design, they used the Pharmacogenomic Research Network Antidepressant Medication Pharmacogenomic Study (PGRN-AMPS), which examined patients treated with citalopram, allowing them to test the study's citalopram model using their algorithm.
The study conducted by Taliaz et al (2021) found that machine learning algorithms could be used to predict individualised antidepressant responses based on combinations of genetic, clinical and demographic factors. The researchers analysed response patterns of patients to three different antidepressant medications in the STAR*D study and used machine learning tools to generate a predictive algorithm. The algorithm was able to produce an average balanced accuracy rate of 72.3% and 70.1% across the different medications in the validation and test sets, respectively. External validation in the PGRN-AMPS produced highly similar results, with a balanced accuracy of 60.5% and 61.3%, respectively. These findings suggest that machine learning algorithms can improve the accuracy of antidepressant prescriptions in large datasets with genetic, clinical and demographic features.
The algorithm developed by Taliaz and colleagues has the potential to be useful for clinical support platforms as part of the emerging field of precision psychiatry. This approach takes into account a person's genetic, environmental and lifestyle variability, all of which impact the efficacy of certain antidepressant treatments. The algorithm can pinpoint differences in these variables, leading to improved psychiatric treatment and prevention outcomes. As more data is accumulated from next-generation sequencing technologies and electronic health records, these tools are expected to significantly improve over time (Taliaz et al, 2021).
The use of machine learning in healthcare is gaining momentum and is proving to be useful in the treatment and prevention of various physical and mental health conditions. With further advancements in technology, more precise tools will be developed to help healthcare professionals identify the most effective treatment options for each patient based on their unique set of characteristics and circumstances. This personalised approach to medicine is expected to lead to better health outcomes for individuals and improve overall public health.