Research Concepts and Opportunities for Early-Career Investigators in Cardiac Surgery

Abstract

Basic, translational or clinic, research is a key component of cardiac surgery. Understanding basic cellular and molecular mechanisms is key to improve patient outcomes and cardiac surgery procedures must be compared with non-surgical alternatives. However, there is limited guidance for early-career investigators interested in cardiac surgery research.

This opinion piece aims at providing basic guidance and principles based on the authors experience.

Research remains critical for cardiac surgery: surgical procedures have valid non-surgical alternatives with which they must be rigorously compared and understanding basic cellular and molecular mechanisms is key to develop new strategies to improve patient outcomes.

This opinion piece is written by authors who are in different career phases and who have a strong interest in cardiac surgery research, with the goal of providing guidance and key principles to early-career cardiac surgeons and anesthesiologists interested in research; the methods are detailed in the Appendix.

NEED FOR FORMAL TRAINING IN RESEARCH

Most successful physician-scientists have completed formal training programs – those may also offer important opportunities to connect with potential mentors within and outside one’s own institution. The National Institutes of Health (NIH), leading academic institutions and professional organizations such as the Society of Thoracic Surgeons (STS), the American Heart Association (AHA), the Society of Cardiovascular Anesthesiologists (SCA), the American College of Cardiology (ACC) and others offer clinical funding opportunities to physician-scientists. Institutional programs are often available and provide a bridge to extramural funding.

To be competitive, there needs to be a strong commitment to the “physician-scientist” pathway. Programs to train medical students in basic and translational science, such as the Medical Scientist Training Program (MSTP) funded by the NIH, have existed for more than 50 years. However, only 7.1% of graduates of the programs go on to train in any surgical field, and of those who do go on to surgical training, only 4.9% eventually attend cardiothoracic surgery, representing the third lowest percentage of nine surgical specialties evaluated in the recent National MD-PhD Program Outcomes Study.(1)

While the journey may not start in medical school as the classic MD PhD journey, the clinician-scientist needs to formulate a parallel skill set of how to design and implement own studies, analyze data, write manuscripts, and apply for funding.

Summer research scholarships prior to, or during medical school are often the first dedicated full time research activities. Trainees who become clinical researchers generally have participated in more prolonged immersion in research training: as a research fellowship during residency (where dedicated time of usually a year or more is set aside for projects and development of research skills), a mentored training program, or a graduate degree intramurally as an early-career attending.

Programs provide built-in mentorship, idea consultation, and regular meetings with deadlines to keep projects moving forward. Careful construction of the project, a team to guide the early-career researcher through inevitable hurdles, departmental resources, and mentor support are critical to success. Projects can include retrospective cohort studies (most likely), systematic reviews (possible) and small prospective observational or randomized studies (least likely given the length of time necessary). Resources come in the form of research support staff, protected research time, and statistical support. A universal rule is that these limited sources are only provided after an individual displays strong desire and dedication for research.

Based on our experience, the following preliminary steps should be taken in order to be accepted in a formal training program: identifying an area of interest, confirming a gap of knowledge by a thorough literature review, using previous literature to guide methodology, and identifying a suitable mentor who has the bandwidth, willingness and the expertise to guide.

The currency of academic medicine is research productivity demonstrated in publications (99% percent completion of a project results in no credit). Successful publication of simple projects such as meta-research or review articles proves an early-career investigator’s commitment and determination. Research mentors are crucial to help early-career investigators advance, as metrics for research evaluation are different than those used in clinical training. Regardless of the ultimate destination - a continued career pathway towards research or one focused on clinical practice - research training and productivity accelerates promotion at all faculty levels. While challenging in term of time management, balancing clinical practice and academic research generally improve professional outcomes in both fields.

Enrollment in a master’s program in clinical epidemiology or health services research is necessary for foundational research and critical appraisal knowledge, exposure to grant writing, and demonstrating continued commitment to a physician-scientist pathway.

STUDY DESIGNS: OPPORTUNITIES AND CHALLENGES FOR EARLY-CAREER INVESTIGATORS

RANDOMIZED CLINICAL TRIALS

Randomized clinical trials (RCTs), which offer the most unbiased comparisons of treatment groups, are considered the gold standard for evaluating the efficacy of novel therapies and the comparative effectiveness of existing ones.(2) Large, practice informing RCTs are complex, expensive, and generally designed and managed by collaborative groups of national or international experts. While mid-career and senior investigators typically participate in such trials, recently there is an increasing focus on including early-career investigators. Approaching Principal Investigators of an RCT that is ongoing, or even in the design stage, with a request to be involved may lead to unexpected opportunities.

Small, hypothesis-generating single-center RCTs may be very important for either testing hypotheses from observational studies or providing sample/effect size estimates to inform the design of larger RCTs. These small pilot RCTs may represent the ideal entry opportunity for an early-career investigator with interest in clinical research who is supported by mentors with RCTs expertise.

There are, however, a few key points to consider:

• When the sample size is small, randomization per se does not guarantee comparability between groups. Even small imbalances in variables closely associated with the outcome may lead to spurious associations when the number of patients is small. It is important that the inclusion and exclusion criteria are designed to guarantee homogeneity and, in case one or more factors are known to be associated with differences in outcomes, stratified randomization should be considered. Also, a fully adjusted intention-to treat analysis is often appropriate;

• Single center RCTs are generally run by investigators highly invested in their success and are known to largely overestimate treatment effects.(3,4) It is important than the maximum possible level of blinding is used and the standard good practice rules for RCTs are strictly followed;

• Pilot RCTs cannot be powered for important clinical events due to feasibility considerations. Therefore, the choice of the primary outcome is critical. The selected measure should be a surrogate of clinical events whose association with important clinical outcomes is very solid and widely accepted in the community, in order to produce important hypothesis-generating data.

As an alternative, it is often possible to leverage the infrastructure of ongoing large RCTs for ancillary studies. The two options are generally to 1) generate mechanistic data (for example, by using imaging or biomarkers) to explain the effect of an intervention (or the lack of it) or 2) to expand the scope of the parent trial (by adding new patient-centered outcomes or cost evaluations). These ancillary studies can be a good introduction to RCTs for early-career investigators and may be particularly appealing to extramural funding agencies. Finally, participating in the broader clinical trials infrastructure, such as serving on event adjudication committees, data and safety monitoring boards, study sections, and institutional review boards, are opportunities for first exposure to the world of RCTs. Trial groups are generally looking for support and offers to help from early-career investigators are often welcomed (see also Figure 1).

Figure 1.

Suggestions of potential directions for research in cardiac surgery by research type. Parts of the figure were drawn by using pictures designed by Flat Icons from Flaticon. AHRQ, agency for health care research; CABG, coronary artery bypass grafting; CDC, centers for disease control; CMR, cardiac magnetic resonance; CT, computed tomography; DAPT, dual antiplatelet therapy; HCAPHS, hospital consumer assessment of healthcare providers and systems; MAG, multiple arterial grafting; MICS, minimally invasive cardiac surgery; MV, mitral valve; POAF, postoperative atrial fibrillation; QoL, quality of life; SAVR, surgical aortic valve replacement; SDOH, social determinants of health; STS, society of thoracic surgeons; TAAD, type A aortic dissection; TAVR, transcatheter aortic valve replacement; TTFM, transit time flow measurement; VSRR, valve sparing root replacement.

MECHANISTIC STUDIES

Clinical research does not answer the question of the mechanisms of the observed phenomena (the “why” question). Defining mechanisms and pathogenesis is important because it may open unexpected venues and opportunities that go far beyond the initial clinical question. Mechanistic research is performed using pre-clinical models that, in the current era, should include simulation and cell cultures rather than animal models.(5)

There are several areas in cardiac surgery that may benefit from “why-questions” and mechanistic investigations (see also Figure 1).

OBSERVATIONAL RESEARCH

For many reasons - elimination of bias, the integrity, accuracy and completeness of data collection, the clarity and validation of outcomes - RCTs are considered the gold standard for comparative effectiveness evaluations. However, for different reasons (mostly equipoise, deliverability and recruitment issues) many surgical questions are not suitable for RCTs;(6) in fact much of cardiac surgical advancement has come from observational series. Observational research can also provide important proof of concept or preliminary data to be tested in RCTs – studies of large registries and databases are also critical to inform statistical estimates for RCTs design. Observational research is the standard approach for the evaluation of an exposure as opposed to a treatment. In cardiac surgery, that could be for example the effect of different patient level characteristics such as age, sex, diabetes, left ventricular (LV) dysfunction and outcome following cardiac surgical interventions. Gaudino and colleagues recently reported that women have persistently higher risk for death and major morbidity following isolated coronary artery bypass grafting (CABG) using data from the STS Adult Cardiac Surgery Database (STS-ACSD).(7) Data from RCTs can be used for secondary observational studies including addressing questions of exposure rather than the study intervention itself, for example, the potential risk of endoscopic vein harvesting on vein graft patency based on data from RCTs comparing on vs off pump CABG. Observational research is also essential for the initial clinical findings such as the association of invasive Mycobacterium chimaera infection and cardiopulmonary bypass (CPB) with the use of certain common heater-cooler units or high dose tranexamic acid and seizures following cardiac surgery.

However, single center data may reflect the particular expertise, demographics or system of care of that center, which may not be broadly applicable. Multi-center studies broaden the scope of evaluation and better assess institutional variation and generalizability. Good examples of collaborations include the regional Northern New England Cardiovascular Disease Study Group and Virginia Cardiac Surgery Quality Initiative and the International Registry of Acute Aortic Dissection. Large registries such as the STS-ACSD, Medicare, and National Readmissions Database provide nationwide “real world” data particularly trends in utilization. With additional programming, merging of datasets or data linkages opens vast potential not intrinsic to any single source including longitudinal outcomes. Census data may open the opportunity to gain greater insight into the socioeconomic factors associated with surgical outcomes.

Larger sample sizes available through observational databases permit well powered assessment of death, or other hard clinical outcomes rather than rely on surrogate outcomes (see also Figure 1). A recently published study using this approach with Medicare data compared the extent of saphenous vein graft utilization for CABG including over 1 million patients with follow-up of over 15-years.(8)

An important limitation is that large administrative databases lack clinical nuance and coding definitions may deviate from clinically relevant definitions. Quality and granularity of the captured data in such databases can vary and, when low, jeopardize attempts at risk adjustment and statistical mitigation of biases. In addition, comparisons of treatments using observational registries can be compromised by selection bias, while metachronous comparisons always involve time biases. Specific databases have additional intrinsic limitations; for example, the National Inpatient Sample (NIS) database includes a 20-percent stratified sample of all discharges from US community hospitals that varies each year making the evaluation of temporal trends problematic. Rigorous statistical methodologies have been developed to attempt to adjust for these effects with application of a growing list of techniques such as regression analyses, propensity score matching or weighting, competing risk analyses, time varying variables, interrupted time series, longitudinal mixed modeling and innovative approaches such as Mendelian randomization.

Provided in more detail later, the rapidly evolving field of machine learning provides opportunities for observational data from existing databases and patient electronic health records using a “data driven” rather than “hypothesis driven” framework, and offers novel and unique insights not previously appreciated. Implicit in these approaches is however the limitation that only known variables can be adjusted for. However, further analytic techniques (e.g., E-value, instrumental variable) can further help to define unmeasured potential confounders.

The “Strengthening the Reporting of Observational studies in Epidemiology” guidelines provide a meaningful framework for adhering to the highest standards in the conduct of observational research.(9) Ultimately, it is the careful conduct and critical evaluation of both observational studies and RCTs that provide the opportunities to advance surgical knowledge.

For early-career investigators, observational research in the form of an institutional case series or case report, is often the first opportunity for a first author publication. Such research projects are well-suited to the early-career investigator as they are feasible using existing data, can be accomplished without substantial medical writing experience, typically only involve a focused review of the literature, and only require descriptive analytics and not advanced scientific training. The time frame from project design to manuscript submission usually is feasible concurrent with clinical or education duties, and without dedicated research time. The entire process from study design through to submission, and revision has positive carryover effects for the next project, regardless of the content.

META-ANALYSES

In the last decades, the volume of scientific literature has expanded exponentially with an average of two new papers published on PubMed every minute. In front of this overwhelming load of data, systematic reviews and meta-analyses represent an appealing tool to summarize existing evidence qualitatively and quantitatively. Indeed, the number of meta-analyses published on cardiac surgical topics increased by >5500% between 1985 and 2020.(10)

There are two main meta-analytic approaches: study-level and individual patient-level data meta-analysis (IPDMA). The former is accessible, quick and captivating especially for early-career investigators because it is based on published, aggregated data that is extracted easily and analyzable by freely available software. Therefore, study-level meta-analyses are appealing first projects in the clinical research realm for early-career investigators with limited or no previous research experience. However, high methodological standards and novelty are key to success for meta-analyses. Early-career investigators contemplating a meta-analysis should identify a research question that is Feasible, Interesting, Novel, Ethical, and Relevant (FINER). Meta-analyses should not be performed if previously published randomized evidence has already consistently showed the presence or absence of a treatment effect, suggesting that there is no unsettled evidence gap. In contrast, a meta-analysis should not be pursued when despite a knowledge gap existing, studies hitherto published are limited or contain substantial clinical or statistical heterogeneity that renders pooled estimates not clinically meaningful.(10)

An example of appropriate meta-analytic application is aggregating clinical outcomes from studies whose analyses were only powered for a primary composite endpoint and underpowered for individual outcomes or safety outcomes. In this situation, pooling data from various studies increases statistical power and can generate novel and potentially important data. For instance, in a recent meta-analysis comparing the use of aspirin or dual antiplatelet therapy after CABG for graft failure prevention, the included randomized trials were not powered for the safety outcome of bleeding, which was a rare event in each study.(11) Pooling the bleeding events from all trials allowed to achieve a higher statical power and provide a more solid answer regarding the safety of dual antiplatelet therapy after CABG.

IPDMAs rely on all patient-level data collected in individual studies and require the collaboration and coordination of multiple research teams. Confidentiality and protection of patients’ data are key priority in IPDMAs, and it may be difficult for early-career investigators to access primary data from these sources. Funding agencies are proactively helping overcome this obstacle. The recent “2023 NIH Data Management and Sharing Policy” by the NIH requires raw data from any NIH-funded studies to be shared, paving the way to an easier access and management of subject-level data. Despite being more intricate, IPDMAs have important strengths compared to study-level meta-analyses. Outcomes such as quality of life scores are better analyzed at patient-level, as often different scales employed are difficult to homogenize when aggregating data. Moreover, IPDMAs allow more thorough exploration of treatment effect modifications exerted by patient characteristics in clinically relevant subgroups, and avoid ecological fallacy.

Early-career investigators approaching meta-analyses should strictly adhere to the methodological process of meta-analyses to guarantee the production of high-quality evidence and should refer to published guides for each step of this process (Figure 1).(10,12)

SPECIFIC AREAS OF INVESTIGATION: OPPORTUNITIES AND CHALLENGES FOR EARLY-CAREER INVESTIGATORS

CARDIOVASCULAR IMAGING

Cardiac surgery relies on imaging to inform patient selection, therapeutic strategies, and post-operative surveillance following an array of interventions. Current-day cardiovascular imaging is predicated on assessments of cardiac size, shape, contractile function (LV chamber size and ejection fraction are encompassed in consensus guidelines for coronary revascularization)(13) valvular heart disease,(14) and linear cutoffs for aortic size which are a cornerstone of guidelines for aortic surgery.(15) In parallel, imaging has been used to differentiate infarcted from salvageable (“viable”) myocardium using conventional approaches such as radionuclide (single photon emission tomography) or echocardiography with adjunctive dobutamine infusion. However, it is well-known that surgical outcomes can vary, including some cases in which patients that meet imaging criteria for surgery fail to demonstrate expected post-operative improvements. For example, among patients undergoing CABG in the Surgical Treatment for Ischemic Heart Failure trial, viability imaging and/or dobutamine echo failed to predict LV functional recovery or clinical prognosis.

Conventional imaging has shown to be limited for patients with aortic aneurysms, as evidenced by data showing that dissection can occur in patients not meeting size-based thresholds for aortic surgery, thus impeding risk-appropriate decision-making for aortic graft surgery as well as risk-appropriate post-operative surveillance.

Taken together, current data in an array of surgical settings illustrate the limits of conventional imaging and highlights need for new approaches to inform surgical decision-making.

Advances in cardiac imaging have enabled new insights of particular relevance to cardiac surgery. Cardiac magnetic resonance (CMR) enables volumetric assessment of cardiac structure and function and (4-dimensional) flow quantification, as well as tissue characterization of an array of myocardial substrates (infarction, non-ischemic fibrosis, ischemia, edema/inflammation). Positron emission tomography (PET) enables assessment of myocardial blood flow and metabolism, including glucose and fatty acid utilization. New advances in echocardiography, including 3D volumetric imaging and strain-based assessments of cardiovascular deformation, respectively enable improved quantification of LV size myocardial and vessel wall biomechanics. Whereas initial studies have suggested these advanced methods to provide incremental utility to conventional imaging, their prognostic utility in has yet to be prospectively tested in large-scale multicenter surgical cohorts.

Key knowledge gaps concerning imaging for cardiac surgery relate to questions on the incremental utility of myocardial tissue characterization (compared to conventional metrics) for guiding intraoperative decision-making and post-operative outcomes. Regarding CABG, it is unknown whether viability assessment using established (bright blood) or emerging (dark blood) late gadolinium enhancement CMR methods provides a robust means to predict surgical response, and whether likelihood of functional recovery is modified by factors such as impaired perfusion or non-ischemic fibrosis on CMR.

With respect to degenerative mitral regurgitation (MR), it has been shown that patients with mitral prolapse can have both sub-annular fibrosis (on CMR) and inflammation (on PET), but it is unknown whether these alterations in tissue substrate predict post-surgical remodeling response or if long-term arrhythmic response persists even after surgical resolution of MR.

Regarding functional MR, CMR has been used to demonstrate strong associations between regurgitant severity and both ischemia and infarction in mitral valve adjacent regions, raising the question of whether patients with sub-valvular ischemia (i.e., minimal or absent infarction) can be effectively treated via surgical revascularization alone (rather than valve repair or replacement). More broadly, given that CMR has been shown to provide a precise tool to quantify valve stenosis and regurgitation and that discordances between modalities have been shown to be frequent, further research is warranted to further test relative utility of CMR and echo-derived diagnostic criteria for valve surgery, as well as whether imaging criteria for surgery should be adjusted to account for factors such as sex, or modified based on factors such as 3D chamber volumes and/or myocardial strain.

Regarding aortic surgery, it is known that new imaging methods to characterize flow and vessel wall properties can differentiate between diagnostic conditions and distinguish native and grafted aortic territories, but it is unknown if differential flow and tissue properties predict long-term adverse aortic remodeling and dissection risk after graft replacement surgery.

Given the growing capabilities of surgery and non-invasive imaging, further research to address these knowledge gaps are of substantial importance (Figure 1).

CARDIAC ANESTHESIOLOGY

Cardiac anesthesia has evolved rapidly, in parallel with cardiac surgery. Advancements in monitoring such as perioperative echocardiography have dramatically improved the anesthesiologist’s ability to diagnose and intervene for surgical planning as well as hemodynamic optimization.(16) The role of the cardiac anesthesiologist has become increasingly important as patients presenting for surgery are older, have more co-morbidities, and are cared for outside of traditional cardiac operating room settings (e.g. cardiac catheterization and electrophysiology laboratories).

Anesthesiology continues to lag behind surgery and other medical fields in research funding from the NIH, and only 3.5% of anesthesia funding is cardiac-related.(17) However, there is a need for a new generation of anesthesiologist-scientists to provide much needed evidence-guided clinical decision-making in the cardiac operating rooms.

Cardiac anesthesia is at the intersection of cardiothoracic surgery, cardiology, and anesthesia, and there are major opportunities for collaboration to answer important clinical questions. The intraoperative period is dynamic and interactive, and provides opportunities to intervene on real-time findings. Unique research opportunities include the use of advanced imaging and intraoperative monitoring to understand, predict and improve surgical outcomes. For example, intraoperative echo strain can detect acute changes in aortic physiology after surgical graft placement that may impact future remodeling and dissection risk in the distal aorta. Three-dimensional changes in the mitral valve annulus may help determine therapeutic response to mitral valve repair for degenerative MR. Assessment of LV longitudinal myocardial strain in combination with evaluation of CABG graft flow may be able to determine the mechanism behind differential early outcomes between patients receiving arterial and venous grafts for CABG. More research on the mechanisms and association of intraoperative findings to outcomes is needed. The next step includes testing if interventions that improve mechanisms underpinning clinical events lead to improved patient outcomes. In order to answer these questions, a new level of collaboration is needed. Intraoperative interventions, often isolated to anesthesia databases, should be linked to surgical outcomes. While the STS has “added on” an anesthesia section to the STS database, more work needs to be done to truly integrate the data.

RESEARCH IN GENDER DISPARITIES

Cardiothoracic surgery and anesthesiology are male-dominated fields with women making up less than 20% of practicing surgeons and one third of practicing anesthesiologists in the US.(18–20) Similarly, the cardiac surgical patient population is overwhelmingly men, with women representing less than one third of the patient population in cardiothoracic surgical trials.(21–24)

Gender disparities research is a powerful tool to identify current and past inequities. This research can focus on disparities among practicing physicians, or in outcomes among cardiac surgical patient populations. The first describes gender gap in academic leadership: the underrepresentation of women as speakers and chairs at scientific meetings, members in clinical practice guideline committees, editors and editorial board members, and as lead and senior authors. While these studies often involve simple descriptive and trend statistics, the “denominator” used to calculate the gender gap should be clearly defined. Depending on the population of interest, the appropriate denominator may be the number of residents, fellows, practicing physicians, or in academic leadership positions, specific to the region.

Gender disparities research also describes differences in outcomes of cardiac surgical interventions.(25) Women have been underrepresented, evidenced by low participation to prevalence ratios (i.e., the proportion of women among trial participants relative to the proportion of women among the disease population studied). Adequate participation of women in clinical trials is important to examine potential sex differences in treatment effects. In addition, cardiac surgery trials should aim to also report women enrollment, sex-specific screening failures and sex-specific loss to follow-up to better understand barrier to women trial participation.

Large scale databases can be used to compare clinical outcomes between women and men. While databases are easily accessible and have large sample size and power, one should be cautious in using administrative, non-clinical databases to answer a clinical question. These databases, such as the NIS or the Agency for Health Care Research and Quality Healthcare Cost and Utilization project lack important variables that affect surgical outcomes such as body size, baseline cardiac function, CPB and aortic cross clamp times. More granular clinical databases such as the STS database and the Multicenter Perioperative Outcomes Group database require formal applications, an approval process, and a cost of $5,000–15,000 depending on the size of the data request. An experienced statistician is crucial for model creation to account for unmeasured confounding on the patient, provider and institutional levels. Time, resources, the quality of the database, and available clinical variables should be considered at the outset. An administrative database may be sufficient to generate a new hypothesis on women and outcomes, but a clinical database may be better to test the previous hypotheses in more detail.

The third category of research on gender disparities is qualitative research to understand patients’ beliefs, experiences, attitudes, behavior, and helps answer the “why” of gender inequity. These can be surveys, questionnaires, or interviews. While traditionally considered less rigorous than quantitative research, qualitative research is a powerful tool to understand areas of improvement. Questions should be created thoughtfully, and coding of answers should occur with the help of a qualitative research expert.

Correct language is important in gender disparities research. Many authors use the terms sex and gender interchangeably, but important nuances exist in academic writing. The traditional definitions of gender and sex define sex as a biological (physical and physiological) differences and gender as how a person identifies and exists as a social construct. Bibliometric research such as evaluating the percentage of women authors or Editorial board members by name-searching should use “gender”. Database research including sex as a variable should use “female” or “male” terminology. However, recent convention in the literature has been to refer to a group of biologically female humans as “women” or as “female patients” and not to use “females” alone to describe female patients.

It is crucial to know the previous literature on the subject. While gender disparities are still a “hot topic”, the bar for publication is higher than ever and requires in-depth analysis and increased novelty of the research question. Fortunately, gender disparities research is valued, and there are many funding and publishing opportunities for women interested in academic cardiac surgery. Journals have special issues focused on women. National societies including the Society of Cardiovascular Anesthesiologists and the American Thoracic Society have specialized funding opportunities for women. The American Association for Medical Colleges offers early- and mid-career leadership programs for women. Academic departments offer funding for research support staff for women investigators with young children. In addition, the NIH now offers up to $70,000 for one year support for early-career women investigators who are pregnant or give birth to preserve research continuity and continued career development.

Overall, research in gender disparities remains an important topic from the perspective of the female patient and female physician in cardiothoracic surgery and anesthesia. Individuals interested in this field should push the boundary of what has been done before and focus on how to optimize women as patients as physicians utilizing bibliometric and database studies, and develop prospective trials on women. Key tips for gender research in cardiac surgery are summarized in Figure 2.

Figure 2.

Key tips for gender research in cardiac surgery. Parts of the figure were drawn by using pictures designed by vecteezy.com and by Freepik and Dimitry Miroliubov from Flaticon.

MACHINE LEARNING

Machine learning has enabled extraordinary realities. Our phones are full of it. Cars can drive themselves. Even some medical devices leverage the fast, efficient flow of machine learning algorithms. It is entirely reasonable to ask how surgical researchers can use these powerful new tools. As a corollary to that, to use machine learning in the literature, one must also need to know how to assess it rigorously and be able to interrogate its assumptions in the same way we do for a RCT, a sampled survey, or even something like a more traditional statistical model (e.g., a linear regression).

It is even a little bit hard to say exactly what does, and does not, qualify as “machine learning.” The old heuristic was something like “machine learning is just fancier, better algorithms for prediction.” But that is no longer true; there are machine learning approaches to causal inference. At the moment, the best way to identify a machine learning approach is to note that the algorithm is being justified through its ability to optimize the prediction of a target variable (e.g., “this algorithm had the best mean square error of all the algorithms we evaluated”; “this parameterization optimized the F₁ score”). This is different than classical statistical models which tend to justify themselves by the plausibility of the model’s ability to represent the real-world (e.g., “recognizing that the response changes non-linearly in age, we included a squared term in age in order to meet the linearity assumption of linear regression”). Early investigator interested in the core of distinction between machine learning and classic statistical modeling then should read the Common Task Framework (section 6.1).(26) The distinction, although ostensibly technical in nature, actually has enormous implications for clinical research. Simply put, rather than “hypothesis” driven research, machine learning, powered by ever-increasing computational processing power, may be “data driven.” In other words, associations identified by the learning algorithms may identify novel associations with unknown explanations. The challenge comes not only in forging a hypothesis, but in identifying the elements that contribute to the association, which may be obscured by the subtle nature of the modeling.

This profound reversal does not permit abandonment of the rules of logic or sound research practice. There has not been any change to the fundamental challenges of rigorous causal inference or rigorous prediction. Machine learning has provided tools that make engaging these challenges easier (e.g., ridge regression is a slick way of doing variable selection) but they do not overcome the familiar fundamental challenges (e.g., if surgeons are sorting patients into the new, experimental intervention on perceived “frailty” but no variable tracking frailty is available in the analyst’s data set then a causal analysis will still suffer from “omitted variable bias” regardless of which whiz-bang machine learning algorithm was utilized). Investigators considering machine learning approaches should be confident that existing intuitions about statistical analyses still hold. They should ask the questions: are the authors making causal claims? If so, it is entirely appropriate to ask about confounding. Are you concerned that an important sub-population of interest (e.g., CABG patients in the age range of 40–50) are not adequately represented in the data set? Then the authors should discuss their data set (e.g., inclusion/exclusion information, how the data were collected).

The most common application of machine learning techniques in cardiac surgery has been in development of risk prediction models.(27) However, the powerful predictive potential of these methods has opened new avenues for diagnosis(28) and therapeutic selection.(29)

In short, in the embrace of this powerful and evolving methodology and its application to research, do not let the uncertain nature of prediction derivation to being unnerving and cause abandonment of the approach altogether; however, do not let its intriguingly mysterious nature to be so intriguing that the fundamentals of rigorous and logical research are abandoned in support of its findings. The rational integration of machine learning will indeed greatly enhance the scope and finding of the surgical research of the future.

An excellent entry point into rigorous machine learning is provided.(30) A list of interesting potential applications of machine learning for future research (suggestive but far from exhaustive) is reported in Figure 1.

Abbreviations:

CABG

coronary artery bypass grafting

CMR

cardiac magnetic resonance

CPB

cardiopulmonary bypass

IPDMA

individual patient-level data meta-analysis

LV

left ventricular

MR

mitral regurgitation

MSTP

Medical Scientist Training Program

NIH

National Institutes of Health

NIS

National Inpatient Sample

PET

positron emission tomography

RCT

Randomized clinical trial

STS-ACSD

Society of Thoracic Surgeons Adult Cardiac Surgery Database