Promises and Challenges of Machine Learning in Laboratory Medicine

September 2023 - Vol.12 No. 9 - Page #10
Download Digital Edition Share Save
Category: Chemistry Analyzers

Throughout recent years, the field of laboratory medicine has witnessed a notable integration of artificial intelligence (AI) in clinical practice, marking important steps toward improving the workflow of data analysis and supporting clinical diagnostics and decisions. Machine learning (ML), a sub-field of AI, develops sophisticated models to analyze large datasets, recognize intricate relationships behind high-dimensional data, and unveil hidden and meaningful patterns. The use of ML in clinical laboratories can serve as an advanced new tool to delve deeply into datasets beyond simple statistics.

Furthermore, the adoption of AI and ML models could potentially lead to reduced turnaround times for laboratory tests, more precise diagnosis, personalized approaches to patient treatment and management, and the promise of improved patient outcomes. As AI continues its gradual integration, the benefits are becoming apparent in the realm of laboratory medicine.

Current AI and ML Applications in Laboratory Medicine

The positive influence of AI on laboratory medicine is not a distant ambition. It is a current, tangible reality that holds the potential to revolutionize diagnostics and enhance patient care. Although most AI-based algorithms are still considered to be in the early development stages requiring rigorous evaluation in ongoing studies, the integration of AI into clinical laboratories and patient care is already yielding transformative outcomes across various diagnostic domains. The following are some notable examples that showcase the power of AI already implemented in clinical laboratories for patient care.1

Hematology Morphology Analysis

Advancements in AI-facilitated morphology analysis have marked a significant improvement in hematology practice. Notably, recent developments have seen the implementation of more advanced image processing equipment and models. AI-driven products for hematology morphology typically capture high-quality images and staining properties, extract key features, and then use supervised models or algorithms to identify, classify, and differentiate cells. Results, particularly abnormal ones, are often flagged to undergo further human review before releasing to physicians and patients.

While the currently available instruments are grounded in AI principles, their specific techniques and algorithms may differ. The automation of traditionally manual processes via AI in hematology laboratories not only helps accelerate diagnostics, but it also enhances standardization and efficiency, thereby reducing manual involvement and interventions.

Infectious Disease

Infectious disease management is another area where AI products and tools have been aggressively developed. As an example, ML algorithms have been applied to identify antimicrobial resistance (AMR) in bacteria and fungi using large next-generation sequencing (NGS) datasets. There is software currently available that analyzes NGS data to predict antibiotic resistance in Staphylococcus aureus and Mycobacterium tuberculosis. Along the same lines, larger analyzer systems have imaging applications that use AI algorithms to interpret bacterial growth and identify specific culture characteristics, as well as analyze plate images for methicillin-resistant Staphylococcus aureus (MRSA).

In addition, MALDI-TOF mass spectrometry instruments have been paired with AI models to improve the rapid identification of microorganisms. By analyzing complex spectral data using AI, diagnosis becomes faster and more precise. In the domain of microscopy image analysis, there is another system that utilizes machine vision algorithms to detect malaria in blood smears. Examples of evolving pathogen detection include assays for influenza A+B, which harnesses a neural network-based pattern recognition system to detect influenza A and B viruses. The beauty of this approach lies in its adaptability to ML algorithms, which can be updated to accommodate emerging viral strains in the face of evolving pathogens. 

Cancer Diagnostics

In the ever-evolving landscape of cancer diagnostics, AI has emerged as a transformative force. Several software systems harness AI to render structured insights and interpretation into NGS tests, guiding oncologists to link clinically relevant mutations to actionable treatment decisions, therapeutic avenues, and clinical trial opportunities.

Beyond these, AI has fostered liquid biopsy tests that analyze patterns of circulating DNA, RNA, and/or proteins, seeking to detect multiple cancers early on, even before symptoms manifest. As the integration of AI technologies in cancer diagnostics continues to expand, the potential for rapid, accurate, and personalized patient care becomes even more promising.

At-home Diagnosis

AI technologies are extending beyond laboratory settings and are increasingly influential in the development of at-home diagnostic tools. Recent blood glucose monitoring systems offer smartphone-connected platforms that leverage AI to identify individual blood glucose patterns, issue timely alerts, and provide insights tailored to each patient. Similarly, there is a smartphone-based kidney test that comprises a kit and a smartphone application for the semi-quantitative measurement of albumin and creatinine in urine to report an albumin/creatinine ratio. Its software performs image analysis via computer vision algorithms.

These AI-empowered solutions are illustrative of a burgeoning technology-based shift in healthcare, merging clinical and personal care for enhanced patient outcomes.

Treatment Optimization

The realm of AI in laboratory medicine is not limited to diagnostics; it is making headway in treatment management with examples in the context of diabetes care. These devices integrate ML to analyze data from continuous glucose monitors and insulin pumps as well as self-monitoring blood glucose measurements to provide optimized insulin dosing recommendations. Using AI, systems like this can help in personalizing treatment protocols, using real-time patient data, streamlining diabetes management, and enhancing patient safety and outcomes.

Implementation Challenges in a Complex Landscape

The promise of AI in laboratory medicine is unquestionable, evident by the high interest and substantial resources invested in the field. However, the translation of this potential into real-world implementation in laboratories of all sizes (and budgets) faces significant challenges at every stage, from model development to validation and clinical implementation.

Model Development

Even the most technologically advanced hospital laboratories encounter difficulties in developing custom ML models. A critical factor in model performance is the quality of training data. Employing a large amount of high-quality laboratory-produced data is paramount to avoid a “garbage in, garbage out” scenario. Insufficient or nonrepresentative training data may lead to biased models, leading to misguided conclusions.

Additionally, ML models can inherit data biases present in training data, leading to discriminatory decision making. Reliance on correlations in complex systems can lead to faulty predictions, highlighting the need to address these developmental challenges in order to achieve widespread ML integration in clinical laboratories.

Evaluation

The evaluation of ML models poses challenges comparable to those of other medical products. As mentioned, individual laboratories’ models rely on specific training datasets, which can lead to performance inconsistencies when applied across diverse settings. An example is the Epic Sepsis Model, which exhibited disparate performance in community hospitals and academic health systems with different patient populations and sepsis incidences.2 Thus, the ML model performance across diverse hospital settings hinges on rigorous validation to enhance model generalization.

Using a different example, researchers at Weill Cornell Medicine (WCM) evaluated the generalizability of an ML model that predicts the normalcy of parathyroid hormone-related peptide (PTHrP). The PTHrP is a tumor marker that can mimic certain actions of parathyroid hormone (PTH). Further, PTHrP can stimulate calcium resorption from bone and reabsorption in the kidneys, thereby increasing calcium concentration in the blood. PTHrP is the most common cause of malignancy-related hypercalcemia, and testing for PTHrP can help diagnose maligancy-related hypercalcemia when the cause of elevated calcium is not clear. However, PTHrP testing often is ordered for patients with a low pretest probability. WCM researchers developed a ML model to predict PTHrP results based on patients’ other laboratory test results that are available at the time of PTHrP ordering. This model can improve the utilization of PTHrP testing and reduce the burden of manual data review.

Of note, when transporting the model to external datasets collected from other hospitals, the model experienced performance deterioration, which was fixed by local customization strategies, such as retraining and rebuilding the model with the site-specific dataset.3 This work demonstrated a strategy to generalize and deploy the machine learning model in different hospital settings or for different patient populations. The example of the PTHrP model highlighted the complexity of model validation across multiple sites.

Implementation

Integrating ML models into clinical laboratory workflows comes with challenges that span technical, regulatory, and operational realms.

  1. Novelty and adaptation. The novel nature of ML requires adapting traditional laboratory practices to accommodate its unique processes. Establishing guidelines and best practices is crucial for seamless integration.
  2. Validation complexity. Validating ML models differs from conventional methods, demanding new protocols. The absence of standardized evaluation measures adds complexity to the validation process.
  3. Opaque decision making. The inherent opacity of certain ML models, often referred to as the “black box” problem, poses a multifaceted challenge. Complex models, like deep neural networks, make decisions based on intricate patterns that are difficult to interpret. Additionally, commercial products leveraging proprietary algorithms might not disclose the underlying decision-making processes, further limiting transparency.
  4. Regulatory considerations. Regulatory bodies, such as the FDA, play a crucial role in ensuring the safety and efficacy of ML models in clinical settings. While AI and ML offer unprecedented potential, navigating the regulatory landscape poses challenges. Commercial ML devices must adhere to stringent FDA requirements to guarantee their reliability and safety.

Conclusion

As the field of laboratory medicine undergoes a transformative shift, AI emerges as a powerful tool with the potential to revolutionize diagnostics and patient care. Diverse applications of AI, from blood analysis to remote diagnostics, underscore the breadth of AI’s impact on clinical laboratories.

As these innovations continue to advance, the boundaries of what can be achieved within clinical settings are being redefined, ushering in a new era of precision medicine. While challenges exist, collaborative efforts among medical professionals, researchers, regulatory bodies, and technology developers are key to realizing the full potential of AI in laboratory medicine. By navigating these challenges collectively, laboratory medicine is poised to pioneer a new era of diagnostics, treatment optimization, and personalized patient care, reshaping the future of healthcare.


References

  1. US Food and Drug Administration. Artificial Intelligence and Machine Learning (AI/ML)-Enabled Medical Devices. www.fda.gov/medical-devices/software-medical-device-samd/artificial-intelligence-and-machine-learning-aiml-enabled-medical-devices. Accessed September 1, 2023.
  2. Lyons PG, Hofford MR, Yu SC, et al. Factors Associated With Variability in the Performance of a Proprietary Sepsis Prediction Model Across 9 Networked Hospitals in the US. JAMA Intern Med. 2023;183(6):611-612.
  3. Yang HS, Pan W, Wang Y, et al. Generalizability of a Machine Learning Model for Improving Utilization of Parathyroid Hormone-Related Peptide Testing Across Multiple Clinical Centers. Clin Chem. In press.


He Sarina Yang, PhD, MBBS, DABCC (CC,TC), is associate professor of clinical pathology and laboratory medicine at Weill Cornell Medicine (WCM) in New York. She also is director of clinical chemistry, toxicology, and therapeutic drug monitoring at WCM. Sarina supports the nurturing of future laboratory scientists and encouraging young professionals in the field.

Austin Jin is a student at Great Neck South High School, New York, who has a great interest in laboratory science, medicine, and technology.




Further Reading on AI and Automation in the Lab

Merge AI and Clinical Chemistry
By Melody Boudreaux Nelson, MS, MLS(ASCP)CM
medlabmag.com/article/1975

Ease the Adoption of Microbiology Automation
By Aaron Odegard, MS, MLS(ASCP) CM, SMCM
medlabmag.com/article/1983

Benefit from Automation and AI in Microbiology
By Jacqueline Getty, MLS(ASCP)
medlabmag.com/article/1853

WHERE TO FIND
Chemistry Analyzers
For a full list of vendors offering Chemistry Analyzers, go to: www.medlabmag.com/findit
Key Chemistry Analyzers Suppliers
Beckman Coulter Diagnostics
Diasorin
LGC Clinical Diagnostics, Inc
QuidelOrtho Corporation
Randox Laboratories
Roche Diagnostics
Login

Like what you've read? Please log in or create a free account to enjoy more of what www.medlabmag.com has to offer.

Current Issue