Packing for the journey: essential partners in innovation

If we are going to make progress on the road to innovation, we need to pack some essentials for the trip. We need scientific insights, a commitment to rigorous research methods, and access to educational platforms to share what we are learning. Academia brings these elements. The development of technology, whether a new wearable device, hardware for more precise cardiovascular imaging, or a new machine learning algorithm, is resource-intensive ^{Reference Faiyazuddin, Rahman and Anand1} The financial capital, engineering expertise, and infrastructure required on the road to innovation today extend well beyond what an academic medical centre can provide. For this reason, we need industry. ^{Reference Kelly and Young2} Innovation also requires access to patients. As innovations roll out, they require standardisation, assurance of quality in real-world practice, and practical guidelines for their use. Professional clinicians and the organisations to which they belong bring together the expertise required to ensure that physicians understand how complex innovations are best implemented in practice and what their strengths and weaknesses are. For innovation to succeed in paediatric cardiovascular care, all three groups, such as academia, industry, and professionals, must be involved. The direction of innovation depends on the motivations each group has for innovation, and an examination of their values is enlightening.

The value of an innovation, and therefore the focus of our efforts, depends greatly on perspective and frame of reference. The American Heart Association recognised this in their 2024 scientific statement about value in cardiac imaging. ^{Reference Hanneman, Playford and Dey3} As a society, a procedure’s value is a simple construct: effectiveness divided by costs. The values of medical academia ought to align with the public interest, and value for them will be effectiveness divided by costs. They will most treasure innovations that are highly effective and cost-efficient. Professional societies strive to serve clinicians, and therefore they value clinical quality, the appropriate use of an innovation to produce favourable outcomes, and good service with a minimum of waste. Industry, let’s be frank, is in business for financial profit and is looking for innovations that give them a profit margin—where they see the greatest revenue relative to cost. If these three stakeholders must travel together, then the road to innovation will be found at a convergence of these admittedly divergent sets of values.

Prediction as the convergence point

One way to anticipate where innovation may be going is with the thought experiment in which we imagine the overlapping interests of these stakeholders and locate the sweet spot where innovation makes them all winners. We argue that the most productive academic–industry–professional collaborations in the future of paediatric cardiovascular medicine share the common theme of prediction. Accurate prediction is a holy grail for professionals because it empowers appropriate personalised care and service excellence, cuts waste, and produces better outcomes. ^{Reference Khanna, Maindarkar and Viswanathan4} It appeals to the values of academics because it unlocks the potential for more effective care at lower cost and risk. Capitalistic industry will be driven to produce this appealing product and market it for the handsome profits it will generate.

It is therefore no surprise that all three groups find themselves on the road to developing innovative methods for clinical prediction, and that this immediately brings artificial intelligence to mind. Because of the convergence of interest from academics, professionals, and industry on prediction, artificial intelligence in CHD is in transition from making diagnoses to making predictions (Figure 1) . Although we are still interested in artificial intelligence support to answer diagnostic questions such as who has a disease and what type of disease, the innovative frontier is in making predictions about events and trajectories for individuals with an established diagnosis.

Figure 1.

Prediction at the convergence of values.

Figure 1 Long description

A Venn diagram illustrates the convergence of values among three groups: Academia, Industry, and Professional Societies. The Academia circle emphasizes effective care and low cost. The Industry circle highlights appealing products, sales, and profit. The Professional Societies circle focuses on personalized care, excellent service, and better outcomes. The overlapping areas indicate shared values and goals among these groups. The intersection of all three circles is labeled 'Prediction,' suggesting that successful innovation in pediatric cardiovascular care requires collaboration among these sectors.

Prediction has several flavours: Event prediction addresses what is going to happen, for example, who will get heart failure, experience arrhythmias, have a stroke, or die. Trajectory prediction predicts when these things are likely to happen. Personalised therapeutic prediction answers the key questions for the individual: What response should we expect to treatment and which treatment offers the best chance of success? ^{Reference Allam, Feuerriegel, Rebhan and Krauthammer5}

Early predictive successes and their limitations

Pilot studies of artificial intelligence-supported predictions in heart disease are already encouraging. For example, a recent single-centre study by Al-Droubi et al. evaluated several types of cancer for which chemotherapy may be cardiotoxic. Clinical, echocardiographic, and electrocardiographic features were collected in over 20,000 patients. Neural network and random forest algorithms were built to assess risk for cardiotoxicity. Their models performed very well, with the area under the receiver operating characteristic curve exceeding 0.9. At the time of publication, the authors report that their system is being integrated into the electronic medical record (EMR) to guide referral of highest-risk patients to specialised cardio-oncology services. ^{Reference Al-Droubi, Jahangir and Kochendorfer6} Furthermore, in 2021, Galli and associates reported the use of machine learning to take the step from diagnosis to prediction of successful intervention in heart failure. ^{Reference Galli, Le Rolle and Smiseth7} This was a multicenter study of 193 patients with systolic heart failure who met criteria for treatment with cardiac resynchronisation therapy. They collected 28 clinical, echocardiographic, and electrocardiographic features and implemented an unsupervised learning algorithm, which identified two phenogroups with significantly different prognoses. Supervised learning, using a random forest algorithm, revealed that there were 16 features predictive of response to cardiac resynchronisation therapy, 8 of which were related to right ventricular size and function. Eleven features were predictive of prognosis. The one with the greatest association was the tricuspid annular plane systolic excursion. They showed that machine learning reliably identified clinical and echocardiographic features associated with outcomes in this population, with quite respectable area under the receiver operating characteristic curve of 0.81 for cardiac resynchronisation therapy response and 0.84 for prognosis. Additional evidence for the promise of artificial intelligence-supported clinical prediction comes from large-scale analyses of routinely acquired electrocardiographic data. Raghunath et al. trained deep neural networks with over 1 million 12-lead electrocardiographics obtained from over 250,000 patients over a 34-year period to predict 1-year all-cause mortality. The model achieved an impressive area under the curve (AUC) of 0.88 while also retaining an AUC of 0.85 in predicting 1-year mortality with electrocardiographics interpreted as normal by physicians. In fact, analysis of a blinded survey of cardiologists revealed that many of the features identified on the normal electrocardiographic for event prediction were not apparent to expert reviewers. ^{Reference Raghunath, Ulloa Cerna and Jing8} These findings illustrate the vast potential of artificial intelligence-based models to utilise data from routine diagnostic tools such as electrocardiographics to predict crucial patient-centred outcomes. Similar artificial intelligence-supported predictions in heart disease are also beginning to emerge among paediatric populations where risk stratification tools have often lagged. Lv et al. retrospectively analysed 630 hospitalised children less than 5 years of age with heart failure over an 11-year period. They developed an externally validated extreme gradient boosting machine learning model that utilised the following seven laboratory variables to predict in the hospital mortality. The model exhibited an AUC of 0.916 in the training set, 0.851 in the internal validation set, and 0.846 in an external validation cohort. ^{Reference Lv, Sun, Yuan, Shen, Baheti and Chen9}

These published reports of artificial intelligence-mediated prediction are encouraging (Figure 2), and I have no doubt there will be more good innovative work like that to come. However, current models such as the aforementioned ones are topic-narrow, time-static, and therapeutically impersonal. In other words, they address single predictive questions about single conditions, they do not accept new data and update accordingly to make their predictive accuracy better next year than it was last year, and the application of these algorithms may provide some guidance, but more comprehensive precision management of individual patients remains a work in progress. Success on the road to innovation will be defined by the discovery of solutions to these issues and how effectively we can apply them in patient care. This seems a tall order, but there are already ideas out there on ways to achieve this.

Figure 2.

Evolution of AI in CHD: diagnosis to clinical prediction.

Figure 2 Long description

The diagram is divided into three sections: Diagnosis, Early Prediction, and Dynamic Prediction. The Diagnosis section includes disease detection and phenotype classification. The Early Prediction section covers event prediction and risk scores. The Dynamic Prediction section focuses on medicine based evidence and digital twins. Arrows indicate the progression from one section to the next, showing the flow of AI application in CHD.

Personalised and dynamic prediction

We might tackle the personalisation problem through an approach sometimes called Medicine-Based-Evidence. This flips the more familiar mindset of reliance on Evidence-Based-Medicine (the world of the randomised controlled trial (RCT) in which we derive the best overall practice) towards Medicine-Based-Evidence, the world of big data in which we derive best personalised choices. It aims to promote a more personalised approach to care using big data, including not only individuals’ clinical and image data but also their genome, microbiome, their environment, and demographics, even input from wearables. ^{Reference Horwitz, Hayes-Conroy, Caricchio and Singer10,Reference Horwitz, Charlson and Singer11}

When deep learning algorithms receive these inputs and are in possession of these same inputs from large numbers of other real-world patients along with outcomes in terms of the responses to treatments that they have experienced, we would be poised to predict treatment responses for our own patient of interest. The idea is to integrate these statistical predictions with mechanistic disease models to construct a “digital twin,” which serves as the real-time digital counterpart of a patient. Medicine-based evidence is thereby capable of dynamically modelling the effects of various treatment decisions in the context of an individual’s specific characteristics. ^{Reference Thangaraj, Benson, Oikonomou, Asselbergs and Khera12,Reference Sarani Rad, Bitaraf, Jafarpour and Li13} In silico testing of various therapeutic approaches before subjecting the patient to them ought to reduce the risks of the lesser-informed approach of plunging ahead with in vivo clinical management by trial and error. Concrete examples of clinically relevant digital twins are already emerging. In patients with scar-dependent ventricular tachycardia, patient-specific heart digital twins have been developed using contrast-enhanced cardiac MRI to reconstruct individualised electrophysiologic models. In a prospective study of patients undergoing catheter ablation for ventricular tachycardia, these digital twins were used to simulate arrhythmia circuits and predict optimal ablation targets. When compared with invasive electroanatomical mapping, digital twin-predicted regions demonstrated a significantly higher prevalence of abnormal and prolonged electrograms and accurately identified zones of conduction slowing. Importantly, these findings suggest that digital twins can meaningfully inform procedural planning, illustrating the potential of patient-specific models to guide therapy prior to intervention. ^{Reference Waight, Prakosa and Li14} This may be of particular benefit for patients with CHD, a field where Evidence-Based-Medicine has come up short because RCTs are difficult to conduct and often fail to provide definitive solutions because of a small number of subjects, their clinical complexity, and disease heterogeneity. ^{Reference Harris, Mackie and Dallaire15}

Digital twins will not be the only way to walk the road to innovation. Another exciting vision for clinical trial design is put forward in the recently published review by Rosenthal et al., which promotes a specific new framework they call dynamic deployment. They recognised that traditional clinical trial designs are static by nature and insufficient to take advantage of modern artificial intelligence systems in real-world clinical settings, especially those involving large language models that continuously learn and evolve. Instead of a process to derive and validate artificial intelligence systems, deploy them, and then evaluate their performance after they have been in service for a while. They advocate for dynamic deployment of artificial intelligence models with continuous real-time updates that allow ongoing monitoring for safety, efficacy, and reliability. Real-time validation and feedback will mitigate performance drift and identify emerging hazards in a timely fashion. They emphasise that regulatory and methodological standards for evaluating adaptive artificial intelligence will be necessary to ensure patient safety and scientific rigour. ^{Reference Rosenthal, Beecy and Sabuncu16}

Data infrastructure

All the approaches discussed require enormous volumes of high-quality, curated, valid data from the real world. Collaborative registries will be foundational, ideally adapted to include real-world EMR, genomic, and imaging data. Along with classic clinical endpoints, they would integrate patient-reported outcomes. They would include full life-cycle data to represent the behaviour of heart disease from before birth through infancy and childhood and beyond. Within these registries, RCTs can still be designed when appropriate, especially for prospective evaluation of the performance of artificial intelligence-supported algorithms for care. Even the ultra-rare conditions which we might think impossible to study through big data might be modelled synthetically using artificial intelligence-enabled registries.

Furthermore, data quality is essential. Our big data must be collected and organised to allow artificial intelligence algorithms to access it. ^{Reference Bernardi, Alves, Crepaldi, Yamada, Lima and Rijo17} Its quality must be vetted, because no matter how good the algorithm is and no matter how much data we have, it does not change one of life’s basic truths: poor data leads to poor predictions. ^{Reference Schwabe, Becker, Seyferth, Klaß and Schaeffter18} Registries and collaborative databases already exist that contain portions of the data we need. But for our innovative purposes, we would want more comprehensive input. They would need to accept continuous updates, and they must allow the ongoing use of their data in the training and revision of predictive models. Alternatively, we might build new databases suitable to our purposes.

Consider the recently published work of Tong et al. who enrolled 23,000 consecutive paediatric patients undergoing congenital heart surgery and evaluated five different machine learning algorithms for the task of predicting several common post-operative complications. In their validation cohort, the highest-performing models provided mean areas under the receiver operating characteristic curve of 0.893 for low cardiac output syndrome, 0.963 for renal failure, 0.929 for pneumonia, and 0.970 for deep vein thrombosis. ^{Reference Tong, Du and Chen19} Datasets like this one might be made ongoing for dynamic updates and could conceivably be expanded to permit broader inputs and evaluation of more endpoints of interest.

Barriers to progress

You can already imagine we face some barriers on the road to innovation. Barriers to adopting these strategies in clinical trials are substantial and probably account for our not making more progress so far. These innovative clinical designs require the collection and organisation of big data, raising serious data security and privacy concerns that have yet to be satisfactorily resolved. Informed consent for participation in such research remains a work in progress. There are concerns that we might fail to be sufficiently fair and inclusive of the poor and traditionally underserved and marginalised sociodemographic minorities. For example, predictive algorithms trained on real-world healthcare data can inadvertently encode structural inequities. Obermeyer et al. showed that a population-health algorithm systematically underestimated illness severity in Black patients, despite comparable disease burden, because of disparities embedded in the training data. ^{Reference Obermeyer, Powers, Vogeli and Mullainathan20} These findings highlight the risk that precision approaches developed from non-representative datasets may unequally benefit patients with greater access to care, further contributing to existing inequities rather than minimising them. Ensuring inclusive data collection and transparent model evaluation must therefore be considered integral aspects of predictive innovation. Research that develops personalised precision care derived for only the more fortunate in society ought not to be our approach. Artificial intelligence, we are discovering, uses considerable energy for data processing. Clinical trials of the future have the novel and disturbing potential for broad environmental impact. And the lack of transparency of the methods may result in a failure to accept the results we cannot fully understand.

Responsibilities of our travelling companions

As we consider the somewhat daunting barriers to progress on the road to innovation, it is important to remind ourselves how our companions on the road might be equipped to help us make progress. Our stakeholders are the professional clinicians, and the collaborative organisations that support their work, the academicians who push the frontiers of science and teach others about it, and the entrepreneurs who invest in progress and promote its practical development. I suggest that each of these fellow travellers has duties we must fulfil if we are to make progress.

The entrepreneurs in industry, as we noted at the outset, are by far the best source of the large investments of financial resources needed for these innovations. They will fund development and commit personnel. They will be building algorithms using data they did not collect, stored in databases they did not create, and so should pay for the privilege, thereby assuring data continuity into the future. They are responsible for production, and their products must be clinically useful, continuously updated, and accessible at a reasonable cost.

Professionals must provide clear direction to industry, to assure that the products that are developed are relevant to clinical needs. Professionals must commit to providing data to the registries. Participation will involve some effort but remember, no data, no predictions. We emphasise that the primary responsibility is theirs to safeguard patient privacy. It is reasonable that privacy safeguards should be a non-negotiable precondition to data use. Clinical professionals can assure this. Professionals must learn to use the products, and this also will require time and commitment. Through it all, we must remain collegial and collaborative, supporting these efforts in our professional societies and interest groups.

Academicians are well positioned to study and evaluate model performance. They need to have a broad view of performance to include not simply predictive accuracy but safety, minimisation of bias, and promotion of equity. Their role is also to educate the medical community and publish what they learn about both the power and shortcomings of these algorithms. They are also the dreamers, the ones who will propose the novel approaches to prediction in the future.

Clinical perspective and conclusion

Before we close, I want to put on my clinical professional hat and share with you what keeps me up at night as a provider of care. I have cared for uncounted numbers of patients with complex conditions that threaten both length and quality of life. The more difficult their care, the more data I have collected. I have commonly reached the point at which no amount of information seems to help me identify the best therapeutic option. Wanting to do best by the patient, but not hearing a clear message in the flood of data—that is what keeps me up at night. This is why I hope I am correct that the road to innovation will lead us to more accurate clinical prediction.

The future of paediatric and adult congenital cardiology will not be defined solely by the precision with which we can image the heart or even the excellence of our technical skills to intervene. These things are surely important, but our progress on the road to innovation will be marked by what we can reliably predict. The convergence of intense interest in prediction among academia, industry, and our professional community will move us all forward together on the road to innovation.