ArtHead- / iStock

Machine learning (ML), which uses predictive modeling tools to inform healthcare decision-making, has made its mark on just about every sector of the healthcare industry, from epidemiology to genomics to oncology. As Tony Badrick, PhD, describes in November’s Clinical Chemistry Q&A, the technique has made specific inroads in image recognition, which has implications for the clinical labs. “A technique called deep learning, a kind of artificial neural network, has especially revolutionized the field. In routine pathology, classification of blood cells, urine sediment, or immunofluorescence stains can all be supported with ML,” he summarized.

The technology isn’t perfect. Techniques are vulnerable to security breaches and data manipulation, as well as the “curse of dimensionality,” the struggle to create an accurate model from an overly large training set. This could lead to “overfitting” and possibly a false prediction model, wrote Badrick.

In the Q&A, moderator Badrick queries experts Giuseppe Banfi, MD, Andreas Bietenbeck, Mark A. Cervinski, PhD, Tze Ping Loh, MB BCh BAO and Ken Sikaris, MBBS, BSc, on the ethics of using machine learning, its future applications, and why it’s important for lab practitioners to stay informed about this technology. “If we don’t embrace ML, healthcare and laboratories will continue their current technical progression blind to the digital data that diligently stores the real-world efficacy of our health efforts,” Sikaris said.

Speculation abounds that automation and artificial intelligence could potentially replace clinical labs, according to Loh. “Such statements underscore the danger of being complacent with new technologies. While laboratory practitioners do not have immediate mortal danger, we should join the march of ML early and be in the driver seat or risk being driven by it,” he advised.

To strengthen their knowledge of ML, several of the experts suggested that clinical labs think collaboratively. “Clinicians accessing laboratory databases risk misunderstanding the laboratory data, and laboratorians accessing clinical data risk clinical misunderstandings. These risks are avoided when a team including both clinicians and laboratorians supervise learning from a combined database,” Sikaris said. Similarly, Banfi offered up the idea of linking up with informatics experts to build a system that would lead to the meaningful and safe adoption of ML in clinical chemistry.

Bietenbeck called for the development of a set of guidelines to assess ML applications. Labs generate large volumes of structured, valuable data on patients, yet “more context is required for more advanced ML applications including interpretation of ‘omics’ measurements. To keep its central role, clinical chemistry must change from being a unidirectional data source to becoming a data hub.”

ML is not immune to biases and mistakes, the experts conceded. “The algorithm is only as good as the training and input data. It is impossible to consider all possible permutations in a clinical setting, and there will be mistakes made by the algorithms that require careful safeguards to avoid harm,” Loh offered.

Addressing other ethical concerns, Cervinski noted that ML’s predictive tools could work against the patient in some situations. An example of this is an insurance company that uses certain pieces of information—such as the ML’s ability to predict the likelihood of type 2 diabetes— to deny coverage or increase premiums. “With any investigation concerning the health and privacy of patient data, there are always ethical issues,” he said.

Experts also weighed in on ML’s future applications in the clinical lab. Use of patient-based real-time quality control in the lab or the identification of “wrong blood in tube” errors are two areas of potential growth, according to Cervinski. The technique might also be used to diagnose parameters such as defined thresholds or high numbers of tests, or in helping to validate a large amount of data to define a link to a specific disease or symptom, Banfi said.

The goal is to keep experimenting with ML and its applications, Bietenbeck continued. “The gained experience will prove to be invaluable.”

Read November’s Clinical Chemistry to gain additional insights on the smart adoption of ML tools in clinical labs.