Using machine learning in pediatric cardio-oncology: we have the questions, we need the answers

“O Deep Thought computer,” he said, “the task we have designed you to perform is this. We want you to tell us….” he paused, “The Answer.”

-The Hitchhiker’s Guide to the Galaxy by Douglas Adams.

Since incorporation of anthracycline compounds into the care of patients with cancer, there has been a clear understanding that, along with potentially curative effects, there are also a number of toxic side effects including on the heart. Initial pathology studies demonstrated tissue injury, and technologies such as electrocardiogram and phonocardiography uncovered functional effects [1, 2]. The next logical step in the progression of understanding these cardiotoxic effects was to look at populations of patients, and in pediatrics the Childhood Cancer Survivor Study in the late 2000s elegantly detailed the long-term outcomes of anthracycline and radiotherapy on the heart [3]. In the ensuing decades, there has been an explosion of literature investigating various ways to assess evidence of dysfunction at earlier and earlier time points, with echocardiography being a workhorse in this capacity [4]. The final iteration of this process is the ability to predict individuals who may develop such cardiotoxic effects, particularly since early intervention is proven most effective [5]. For survivors of pediatric cancer, there are useful risk calculators that incorporate patient and treatment factors to determine the risk to develop heart failure, ischemic heart disease, or stroke by age 50 [6, 7]. These were created the old-fashioned way, by groups investigators slogging through data on thousands of patients. However, there are no good risk predictors that make use of the many echocardiograms being collected for this patient population, and likewise none that can predict risk in a shorter time frame such that individuals may be identified while still relatively close to their therapy, or possibly even during treatment.

Artificial intelligence and machine learning are cutting-edge technologies being applied in medicine, including cardio-oncology, in an effort to improve a number of aspects of patient care [8,9,10]. While the specifics of how this works and the ethics of its use are beyond the scope of this editorial, their encroachment into daily life suggest they are here to stay. In simple terms, machine learning can process data volumes that exceed the capabilities of even large groups of researchers, and can additionally identify trends in data that could be imperceptible to even the most astute human observer [11]. But, while many of us see a black box of data in/results out with this technology, in truth it is necessary to “teach” such systems what they are “looking for” and how they should be looking for that answer. In other words, machine learning may be able to provide clinicians with an answer, but it is dependent on those clinicians to ask the question and feed them with the right data.

In the current issue of Cardio-Oncology, Edwards and colleagues set out to build a machine learning-assisted tool to predict children at risk to develop cancer therapy-related cardiomyopathy (CTRC) using clinically collected echocardiographic images as the primary input [12]. To accomplish this, they employed a Deep Convolutional Neural Network (DCCN) that could be trained on different sets of echocardiograms from pediatric patients treated for cancer. Two groups of patients were identified: those who developed cardiomyopathy (case), defined as shortening fraction </=28% or ejection fraction </=50% on 2 studies or on 1 study if remodeling therapies were started, and those who did not develop cardiomyopathy (non-case), defined as SV>/=30% or EF >/=55%. Montages, or sets, of 4 single, still-frame images from the parasternal short axis of an echocardiogram at different periods of the cardiac cycle as guided by ECG tracings were used as input data. In total, 115 at-risk patients were included (70 case, 45 non-case), with 542 pairs of montages from longitudinal echocardiograms. Time points went as far back as 2 years prior to development of cardiomyopathy, and as recent as development of cardiomyopathy. What they report is that the DCCN was able to predict reasonably well those individuals who would go on to develop cardiomyopathy, especially when more echocardiograms were included. As the data became smaller, meaning only the older scans were used, performance was decreased.

Overall, this proof-of-concept is an exciting step toward improved prediction of early development of cardiomyopathy in pediatric patients treated for cancer. If validated in larger studies, the method could be applied to any number of potentially cardiotoxic exposures. There are, however, some questions regarding the method. One of the major limitations is the type and quality of data fed into the algorithm. Machine learning relies on two main factors: “What do I want to know?” and “How do I get the answer?” The latter question will depend on the instructions or the recipe and the ingredients, i.e., the data provided. In this study, 4 low-resolution, static images of a small segment of a dynamic and geometrically complex ventricle may not be sufficient for making a complex prediction, and that can explain its less accurate performance on predicting the outcome when studies were further away from the time of cardiomyopathy diagnosis.

In addition to methodological concerns, there is also the question of the ultimate applicability of this new technology. One is reminded of myocardial strain, which has been available for nearly 3 decades and has countless manuscripts dedicated to use as a tool for early detection of CTRC, but in practice it has yet to find a broad foothold in pediatric practice. Along with methodological concerns, there is also the question of the ultimate applicability of this new technology. For example, despite the promise of advanced imaging techniques, many tools struggle to gain traction in clinical settings due to the complexities of integrating them into existing workflows. This highlights the importance of not only having innovative solutions but also ensuring that they are reproducible and accessible to all institutions. Finally, the educational aspect remains a priority; training healthcare providers to interpret and utilize these new technologies effectively is crucial for their success. Without addressing these challenges, even the most sophisticated algorithms may remain underutilized.

Ultimately, it seems inevitable that artificial intelligence will be a regular part of medical care in the future. While some practitioners fear it spells the demise of the human provider [13], others are excited it will expand the role of the physician [14]. But what seems obvious, at least to the authors of this editorial, is that, like all prior technologies, it may provide answers but the question must still come from human intelligence. More importantly, the interpretation of that answer, how it impacts care, and how it is delivered to the patient seem to all safely be the domain of a human provider. Because sometimes, the answer alone is not enough. Even after 7.5 million years of contemplation, Deep Thought’s answer to the question of life, the universe, and everything was, “Forty-two.”