Heart failure affects over 64 million people worldwide, posing management challenges and significant medical costs. Over the past two decades, there has been a paradigm shift in HF patient care due to goal-directed pharmacotherapies and expanding device therapy options. Machine learning (ML) has also been integrated into HF patient management, allowing for the real-time integration of data and the development of predictive models.
However, challenges remain in the integration of ML into clinical practice, including data standardization, storage, and security, human-AI interactions, and the potential for healthcare inequities. To mitigate these concerns, models applicable to a broader range of populations are needed, along with rigorous testing and continuous learning in the post-marketing phase. In Arkansas, AI has the potential to improve healthcare delivery through the identification and closer monitoring of high-risk individuals. Telehealth and smartphones have also shown the potential in improving healthcare delivery.
According to the European Heart Journal, the integration of ML within existing workflows for the management of HF is promising, but numerous challenges need to be addressed. One of the major issues is the lack of standardized data collection and storage. More than 80% of healthcare data is stored in an unstructured format (e.g., medical images, clinical notes, etc.) and is not often available or utilized in developing ML algorithms. While solutions that employ natural language processing have been devised for data extraction, such methods need further validation before being routinely employed in clinical and research practice.
Moreover, data pre-processing frequently requires manual input, which is then fed into an artificial intelligence (AI) system. Thus, human–AI interactions require a transparent and standardized reporting system. This becomes especially relevant in cardiovascular imaging since image interpretation often necessitates subjective assessment.
Further, with the publication of multiple ML algorithms for the prediction of device therapy endpoints, ensuring uniformity in trial design and model training will be fundamental. This is to guarantee that accurate benchmarks are established so that novel algorithms’ performances can be compared and contrasted with those of existing algorithms. To advance this idea, the SPIRIT-AI guidelines have been introduced to standardize reporting of clinical trial protocols involving ML.
With the need for big data with ML applications, challenges inherent to data storage and privacy need adequate consideration. Solutions such as blockchain technology are being proposed to decentralize data storage while simultaneously augmenting data security. Additionally, with more and more remote monitoring devices being manufactured by third-party companies, formal data usage, and sharing regulations are essential to prevent the unauthorized dispersal of sensitive information.
While the integration of ML within existing workflows for the management of HF is promising, numerous challenges need to be highlighted. For instance, a substantial number of studies involving ML have been performed using retrospectively collected datasets or using carefully curated clinical trial data that do not necessarily reflect real-world knowledge. While results regarding the performance of ML algorithms have generally been consistent, whether these algorithms can translate into actionable insights in real-world scenarios merits further investigation.
The question that follows is whether an ML-structured approach will help with mitigating healthcare inequities. The answer is seemingly complex, with challenges inherent to technology, affordability, adaptability, and applicability needing thorough consideration prior to ensuring wider adoption by the general population. The use of remote monitoring devices is bound to increase the financial costs associated with HF therapies, and not surprisingly, lower socioeconomic and educational status have been associated with lower adoption rates of digital technologies.
With the dawn of the digital era, the increasing adoption of remote sensor technologies will empower patients to become more proactive with their health. But, it will come at the cost of hindering the most vulnerable patients as a result of the need for digital literacy and the socioeconomic support needed for effective application. This might widen the gap in healthcare delivery between the rich and the poor, potentially exacerbating health inequities.
Technology that simplifies patient-level requirements while bearing minimal out-of-pocket costs is imperative to mitigate the above-mentioned concerns. Moreover, models applicable to a broader range of populations that account for demographic and socioeconomic factors, such as race, sex, ethnicity, etc., are needed to avoid biased and inaccurate results. For instance, a lack of diversity in the training dataset risks the generation of a biased ML algorithm based on racial and social backgrounds.
