Implementing AI tools: Best practices and considerations

If you have been to a pathology or diagnostics conference in the last couple of years, you have already encountered it: the seeming whirlwind of AI-powered tools available for clinical laboratory use. It’s a radical change from where this field was just a few years ago. Indeed, the evolution of these tools is happening so quickly that the array of potential options can differ dramatically in a matter of months.

With a rapidly changing landscape, it can be difficult to navigate best practices in AI-powered tools for interpretation and lab results. Which options are worth your time, and which are a waste? Which will lead to better care for patients, and which will end up frustrating your clinical colleagues?

These are critical questions as laboratory teams are bombarded with news and offers about the latest AI services and products. Tools for scanning digital pathology images and suggesting molecular findings, for example, might help staff members review cases faster and deliver results to physicians sooner. The same goes for AI tools designed to accelerate literature review for variant interpretation. But some AI tools could be a burden for clinical teams, or worse, produce inconsistent, incomplete, and therefore untrustworthy results that shouldn’t be incorporated into patient care.

Amid all the uncertainty around AI tools for clinical labs, one thing is clear: demand for molecular testing will only rise in the coming years. With advances in research and new assays being developed and approved, such as minimal residual disease tests that will involve serial testing of the same patients over time, the workload for molecular laboratorians is set to grow substantially. For many labs, AI tools may be the only effective option for meeting that demand in a resource-constrained environment.

Challenges and opportunities

In some ways, these AI tools have arrived at the perfect time. Clinicians are busier than ever. Meanwhile, the reports they’re getting from laboratory medicine practices are growing in complexity thanks to large gene panels, exome or whole genome sequencing, and an increasing number of treatment options that depend on molecular testing to match patients to drugs or to clinical trials. In addition, results are needed faster than ever to help clinical teams make time-sensitive decisions about patient care. Generating results rapidly and then interpreting and presenting them in a clinician friendly manner are essential responsibilities for already overworked lab teams.

Using the power of AI tools to review and summarize the scientific literature, identify drug options associated with a specific biomarker, or scan through a patient’s electronic health record to pull out useful elements could help clinical lab teams get through tedious research faster and focus their time on more valuable tasks. For example, AI is a natural fit for keeping tabs on an exploding knowledge base.

But deploying AI services has to be done in a way that makes sense for both laboratories and patient-facing clinicians. Within the lab, it’s important to match AI tools to the existing workflow, rather than trying to reinvent laboratory processes. Everything from sample processing to how results are communicated should be considered. A critical component to leveraging AI tools is to ensure they fit seamlessly into the workflow. If they don’t, users are less likely to take advantage of them and more likely to view them as a burden rather than a help.

The same is true for ordering physicians. If the process of using an AI service is difficult, clinicians probably won’t adopt it. We already know that many eligible patients don’t get the molecular tests they should have simply because the ordering process is too clunky or inconvenient. If AI tools are implemented in clinical labs in a way that makes the clinician’s workflow more difficult — whether that’s how tests are ordered or how results are reported out — then these tools will not be successful.

An optimal model

When searching for AI services, it can be tempting to look for a one-stop shop to deliver on the clinical laboratories’ strategic priorities: a single solution that can orchestrate a lab’s complete AI needs. Unfortunately, at this stage of AI product development, there is no end-to-end AI workflow. Right now, the best approach is to consider task-specific agents: a tool for reviewing patient records, another tool for summarizing the literature, and so on to build upon the goals of the laboratory’s program. Throughout the process, it is imperative to keep humans in the loop. When it comes to patient care, very few AI tools have been validated extensively enough to go unsupervised by a lab expert. After all, healthcare is still a human-to-human interaction, and we cannot lose sight of this basic fact.

Another avenue to consider, especially for labs with severe resource constraints, is the choice of tool versus partner. With the right team members and sufficient time and money, lab staff could bring in individual AI agents for many of the tasks in a molecular diagnostics workflow. But in labs without those resources, it may make more sense to find a service partner that can recommend the best tools and even help implement them. The partnership model can be even lower cost for labs willing to partner in critical research; many companies can find ways to extract value from insights to advance science and find new ways to treat diseases.

Data curation

AI is the ultimate manifestation of that old adage: garbage in, garbage out. Getting good performance from any AI tool depends on having properly curated, standardized data. This is a huge stumbling block in the healthcare field, where even the cleanest records from two separate laboratories often aren’t concordant with each other. Consider situations where clinicians and some labs use different terms to describe the same thing — such as a position point on a chromosome versus a protein effect — and it’s easy to get a sense of how difficult it is for AI tools to mine electronic health records and emerge with useful results.

For larger laboratories looking to expand their operations into areas such as biomarker discovery or managing clinical trials, it’s important to curate available data properly. If this isn’t something that can feasibly be done by staff members, there are companies that provide data standardization and curation services to help labs get on track. Once the data sources are well organized, any AI tool implemented to help with biomarker discovery or other data mining tasks will be far more effective and reliable.

Whether labs handle data curation themselves or work with a service provider, it’s important to align these processes with how each lab expects to use that data. Standardizing electronic health records is a very different task compared to aggregating oncology test results in preparation for biomarker discovery.

Calling all innovators

Change in medicine is never easy — especially when lives are on the line. But to improve patient outcomes and advance care, we must embrace new ways of thinking. This means overcoming the challenges of transformation and finding ways to seamlessly connect vast amounts of information across the patient journey, enabling better decisions at the point of care. AI products and services are mostly being developed and offered by third-party companies, but it’s the clinical laboratory professionals who have to be the innovators. Only this community can determine the tasks where AI will be most effective, the ideal ways to roll out new AI-powered services, and how to use AI responsibly with sensitive patient data and with end-user providers and other clinical team members who may be skeptical of newer technologies.

With staffing shortages, cost pressures, and overwhelming demand, it can be easy to dismiss AI tools as an unnecessary time sink. But these tools have real potential to help clinical teams increase their efficiency across the patient care continuum, such as translating test results, identifying possible action steps, listing therapies associated with a specific biomarker, suggesting alternatives when patients develop treatment toxicity, and recommending post-treatment management plans.

For all these reasons, AI tools should be considered viable options for clinical laboratories, so long as they are paired with appropriate guardrails to ensure their responsible use. In this way, they are no different from any other novel technology that lab teams evaluate, validate, and implement to enhance their workflows and help their clinician partners deliver even better patient care.