Challenges in the Implementation of Al in Chronic Disease Management

Methodical Challenges

Al’s capacity to enhance and improve healthcare is not without its challenges (Challen et al., 2019). ML model bias and clinical safety concerns specific to chronic healthcare management are referenced and addressed as follows:

■ Most Al models are based on retrospective studies, demonstrate association, and do not prove causation (Thiese, 2014; Yu et al., 2018). Random data variations without a real biological basis may be erroneously classified as a phenotype (Ascent of machine learning in medicine, 2019). The solution for this is methods that incorporate causal discovery algorithms or hybrid approaches which are combinations of probabilistic and causal discovery methods (Glymour et al., 2019)- Further prospective controlled studies are needed to assess these ML models in a real-world clinical setting (Yu and Kohane, 2019)-

■ Models are only as good as the data they are trained on. Even perfect models are restricted by training datasets, dataset quality, and degrees of the signal. Further, most ML studies are trained on databases that comprise a specific population, which harbors the potential for “distribution shift” (Amodei et al., 2016) and may hamper a model’s general applicability to diverse population groups. “Out-of-sample inputs” is

a mismatching condition between training and operational data that may result from either training data inadequacies or incorrect trained model application to an unanticipated patient situation (Amodei et al., 2016). Different outcome definitions are used during the ML model development, which may also cause additional reproducibility problems.

■ Excessive learning by neural networks also creates models that are prone to overfitting, which reduces data applicability to real-world scenarios (Nagpal, 2017). Regularization is a process that is used to reduce over-fitting; however, its role remains to be tested and utilized for model applicability across a wider general population (Nagpal, 2017).

■ Prediction drift is a mismatching phenomenon between training and operational data because of changes in disease patterns over time.

It may affect a model’s accuracy (Davis et al., 2017). Dichotomization is performed in models to classify patients into discrete categories. Inappropriate dichotomization, however, may reduce a model’s predictive precision (Johnson et al., 2018).

■ Unsupervised ML models may perform as a “black box” with unexplained decision-making rationale that is opaque to interpretation (Yu et al., 2018). This lack of transparency in a new technology decreases healthcare providers’ and patients’ trust in such algorithms and may contribute to lower adoption rates.

■ Reporting standardization is not yet established for ML model accuracy (Forman and Scholz, 2010; Lobo et al., 2008).

■ ML models adoption may lead to “automation bias”, which is a phenomenon of clinicians accepting the decision from the algorithms without confirmatory investigations (Parasuraman and Manzey, 2010; Tsai et al., 2003).

■ Concerns regarding the possibility of bias against minorities in underlying datasets, which will be inadvertently learned by the ML models and will influence the output (Romei and Ruggieri, 2013).

Practical Challenges

Al technologies have successfully integrated themselves into healthcare fields such as fundus imaging, chest X-ray interpretation, histopathology, evaluation of skin lesions, ECG monitoring, and hybrid closed-loop insulin pumps to name a few. Al use in chronic disease management is evolving; however, many challenges remain to be addressed prior to its industry-wide adoption:

■ Data Quality - High-quality training dataset availability is currently limited. End-to-end ML model development needs large-scale data collection procedures that require data sharing. Sharing patient health records is highly regulated, and patient data privacy for ML method development remains a point of contention (Yu et al., 2018).

■ Cost - ML model research requires high development costs and expertise. Significant resources are needed to extract and normalize data and customize models to suit organizational requirements. Further, shortterm financial investment benefits have not yet been realized, and only indirect society benefits via potential early diagnoses and improved patient quality of life have been demonstrated. Al researchers have not yet performed large outcome studies that demonstrate improved outcomes and lower healthcare costs, and therefore have not established or determined a quantifiable value to chronic disease management (Rumsfeld et al.. 2016).

■ Clinical Integration - Al models require integration into clinical workflow to demonstrate their full beneficial potential. Al integration, however, is hampered by current EHR infrastructure limitations. Execution of the computational environment for collecting, storing, and sharing confidential health data has its challenges (Yu et al., 2018).

■ Interoperability - Third-party application interoperation with EHR systems is not often successful; however, there are some on the horizon with the development of interoperability frameworks such as

FHIR (Fast Healthcare Interoperability Resources) that can help with Al integration in clinical DSSs (Yu et al., 2018). Framework design standards are also crucial for Al implementation in clinical practice; however, they are not yet available (Middleton et al., 2013).

■ Leadership - Healthcare executives do not currently possess in-depth knowledge about ML model capabilities, challenges, and limitations for chronic disease management.

■ Legal - Current legal and ethical frameworks do not address direct or indirect discrimination, liabilities, and responsibilities to regulate Al management for chronic disease (Romei and Ruggieri, 2013).

Solutions

Al models used for chronic disease management may have unintentional harmful consequences to patient care if their challenges and limitations are not addressed. To address these challenges, healthcare providers, developers, Al researchers, and administrators need to collaborate to develop high-impact Al applications and infrastructure required for successful utilization. Successful Al model development will require correct study design determination, adequate sample sizes, appropriate variable use, and development of complete data protocols. Models must also have abundant and easily accessible data to adequately serve diverse population demographics. Al model validation should be performed ideally through randomized clinical trials in diverse settings to ensure result reproducibility and model generalizability to various clinical settings and patient population. Researchers should ensure diversity in age, ethnicity, and gender in model validation and training datasets to develop widely applicable models. Deployed models must also be continuously refined and trained with realtime patient data.

Successful Al model utilization requires industry adoption and regulatory infrastructure development. Physicians should champion Al adoption, serve as subject medical experts to healthcare systems, and play leading roles in model development, validation, and implementation. Regulatory agencies should establish diagnostic standards and legal and ethical frameworks to ensure comprehensive model performance and security evaluation in real-life clinical situations. ML model regulations have already initiated with the FDA’s “pre-certification” announcement in April 2018, which will intend to give certification to the technology developer rather than the product itself (Transforming FDA’s Approach to Digital Health, 2019).