Artificial intelligence: is it the right time for clinical laboratories?

The term Artificial Intelligence (AI) was originally used by John McCarthy at the Dartmouth conference in 1956 and was defined as the “theory and development of computer systems which perform tasks that normally require human intelligence” [1]. Since then, periods of reduced funding and interest, called “AI winter”, have been followed by significant theoretical and practical advancements [2]. The last two decades have brought significant improvements in mathematics, computer hardware and software programs, and AI began to grow rapidly in different fields. In healthcare, the explosion of digitalized data opened up opportunities for clinicians to use AI, offering them unprecedented opportunities to improve patient outcomes, reduce costs, and personalize care [2]. As a central part of the healthcare system, clinical laboratories have been coping with incremental improvements in informatics for decades, and have been pioneers in digitization and computer-assisted tools as software. This fact, in addition to the central role of clinical labs in patient healthcare, highlights the importance of improving timely and accurate diagnosis and patient care through AI [3]. As a result, the laboratory medical profession may now be facing a big transformation due to disruptive technologies, namely digitalization, Big Data, AI and machine learning (ML).

This special issue entitled “Artificial Intelligence and Big Data in Laboratory Medicine” offers readers the opportunity to get in touch with professionals working in this field (data scientists, but also specialists in laboratory medicine), and discover some of the studies that are based – in part or in whole – on laboratory results. To begin, we asked two philosophers to explain the connection between philosophy and AI, given the role philosophy has played in the development of AI. The key message of the paper by Plebani and San Mauro is that reflecting on the connection between philosophy and AI helps to realize that philosophical and scientific progress are interconnected [4]. The paper by Pennestrì and Banfi introduced the issues of ethics and law, both of which are relevant for data collection and sharing [5]. The authors underline that ethics should not be reduced only to compliance with rules and that too complicated rules could represent a wrong approach, as they could be easily neglected or violated. On the contrary, few clear rules with simplification, flexibility, and professional AI training, are probably a better approach to guide the ethical application of AI in pathology and laboratory medicine [5]. The paper by Padoan and Plebani emphasizes that clinical laboratories generate tons of data every day, often with most of the characteristics of Big Data [6]. Importantly, laboratories generate daily not only test results (analytical phase), but also data from pre-analytical and post-analytical phases, which are continuously generated at all stages of the “brain-to-brain” loop [7], data that are only partially captured in LIS [6]. The paper by Negrini et al. focuses on the applications of AI during the COVID-19 pandemic. Containment of the spread of SARS-CoV-2 prompted many researchers to develop predictive models for rapid and accurate identification of patients with COVID-19 or at higher risk of adverse disease progression [8]. They performed a PubMed literature search and identified 678 articles on AI and COVID-19 topics; of these, 292 were analyzed, and the results underscore that most studies included imaging data alone or in combination with laboratory results. Among laboratory tests, C-reactive protein (CRP) and white blood cells (WBC) were the most commonly used, concluding that, laboratory medicine is gaining momentum, especially with digital tools characterized by low cost and widespread applicability [8]. In a similar analysis, Carobene et al. evaluated 279 studies identified through a database search; of these, 34 were diagnostic and 34 prognostic studies [9]. Interestingly, the authors subdivided the studies according to the country of origin, number of subjects, data collection periods, type of patients and purpose of the tasks. They concluded that the research projects are characterized by extraordinary heterogeneity in terms of patient types, that regardless of the study type laboratory tests should be well described and characterized, and that effective collaboration data scientists is essential [9].

Several papers in this issue report original research in the field of AI predictive models to improve available formulas for computed tests or to predict outcomes. For example, Topcu et al. evaluated a total of 183 ML models for predicting urine osmolarity, using data from 300 spot urine samples, that included urinary conductivity results [10]. The aim of this study was to develop a novel formula for estimating urine osmolality using cutting-edge ML tools to improve existing estimation methods. The results underline that urine osmolality can be estimated using ML models that utilize conductivity and other urinalysis parameters, achieving better results than existing formulas (the ML R² score was 0.83) [10]. In the study of Kurstjens et al. laboratory measurements from 3,797 anemic primary care patients were used to develop and validate an ML algorithm for predicting low ferritin levels. This algorithm is a valuable diagnostic tool that can assist physicians and specialists in laboratory medicine, but can also automatically identify unrecognized iron deficiencies with good performance [11]. Another study by Constantinescu et al. investigated the possibility of integrating ML into the interpretation in mass spectrometry steroid profiles, and described an approach based on some algorithms that return the probabilities of unilateral primary aldosteronism, to facilitate and improve diagnostic decision-making [12].

Other studies published in this issue evaluated the utility of ML for predicting ascites grades or variceal grades in patients with cirrhosis, and showed relevant results for classification accuracy and the algorithms precision [13], [14], [15]. Pei et al. reviewed the use of AI in clinical applications for the diagnosis, treatment and prognosis of lung cancer, and emphasized that in some diseases integration of data (including molecular biomarkers, imaging data and histopathology results) is important to improve the accuracy of predicting disease risk and response to treatment, and is therefore also relevant to the use of AI [16].

Errors in laboratory medicine are now widely recognized as potential diagnostic errors that can have harmful consequences for patients [17]. Delta check (DC) is widely used to detect sample misidentification and wrong blood in tubes (WBIT), and studies have shown that the accuracy of these rules ranges from 15 to 76% [18]. In the paper by Zhou et al., hematology test results were used to evaluate the performance of six ML approaches for detecting sample mix-ups, which showed high accuracy, ranging from 85.5 to 93.1% [18]. On a similar topic, the study by Farrel et al. investigated the effect of human supervision on an artificial neural network trained to identify WBIT errors. The results clearly showed that models classified WBIT errors with higher accuracy than laboratory staff (92.5–81.2%), ultimately confirming the utility of AI in detecting this type of pre-analytical error. Another topic of interest in error detection is the utility of patient-based real-time quality control (PBRTQC), which has recently gained attention as an alternative/integrative tool for internal quality control (IQC). The study by Zhou et al. showed that ML algorithms can be efficiently used as IQC strategies and that they could outperform classical moving average-based PBRTQC by reducing the number of out-of-control patient samples from 600 to 20 (in worse scenarios) [19].

Finally, the question whether laboratory specialists are ready for the digital transformation remains unanswered: the potential application of ML models to laboratory data could be relevant, but to manage the change and uncover additional benefits to patient care, there is an urgent need to adapt expertise within laboratories and to improve the cooperation between laboratories and AI experts. However, as Bellini et al. underlined, some basic prerequisites for the digitization process seem to be still missing, while some infrastructural barriers exist, such as the integration of healthcare data, and the insufficient speed and quantity of data extraction from LIS [20]. Advances in the understanding of biology, pathophysiology of diseases and molecular medicine combined with technological developments have given laboratory medicine a central role, ranging from maintaining well-being to disease prevention, early detection, prognosis, monitoring and guiding personalized therapy. It is time to more appropriately manage the interpretation and use of the huge amount of data generated daily by clinical laboratories to improve patient care, and AI is a fundamental tool to achieve this goal. First and foremost, however, clinical laboratories must ensure that laboratory data are accurate and reliable to avoid the risk of sophisticated systems such as ML and AI using inaccurate results which in turn can lead to inaccurate and potentially harmful information. This concept is well summarized in the mantra “garbage in, garbage out”.