The role of AI in pre-analytical phase – use cases

Research applications

The detection of clotted specimens represents a critical quality control measure in laboratory medicine, as clotted samples can lead to inaccurate coagulation test results and potential analytical errors. Two distinct studies demonstrate the potential of AI in clot detection (Table 1). Fang et al. employed a neural network to identify clotted specimens in coagulation testing based on laboratory parameters such as thrombin time, fibrinogen, partial thromboplastin time, prothrombin time, and D-dimer [19]. The model showed excellent performance (AUC 0.97, accuracy 95.3 %, specificity 96.7 %, sensitivity 94.0 %), with logistic regression coefficients provided for partial interpretability, though noted limitations regarding age disparities in the data [19]. Complementarily, Hou et al. developed an improved UNeXt segmentation network to detect clots and fibrins in serum images, addressing the operational risk of sampling needle blockage [20]. Trained on 9,500 serum sample images and validated on 13,230 real-world cases, the model demonstrated sensitivity of 95.74 %, specificity of 98.11 %, and accuracy of 97.93 % [20].

Commercial solutions

Several commercial products, such as the Abbott i2000SR, its successor the ci8200, the Roche cobas 8100, and the Tecan PMP, incorporate clot detection technology to ensure sample integrity (Table 2). These systems use different methods to detect clots: the Abbott models use probe-based detection, the Roche cobas 8100 employs module-based multi-parameter checks, and the Tecan PMP utilizes pressure-profile monitoring. However, none of these manufacturers explicitly state that their systems use AI or ML for clot detection.

Research applications

Wrong blood in tube (WBIT) errors represent one of the most critical preanalytical challenges in laboratory medicine, with potentially severe clinical consequences for patient safety. Numerous studies have investigated ML models for detecting WBIT errors (Table 1).

Decision tree-based ML algorithms showed promise in detecting WBIT errors. Farrell et al. developed an XGBoost model using CBC data [21]. When applied to retrospective test data, the model demonstrated accuracy of 98.2 %, sensitivity of 96.8 %, and specificity of 99.6 %. In prospective evaluation of 38,187 routine CBC results over 22 weeks, the model detected 12 WBIT errors missed by routine checks [21]. Likewise, Mitani et al. developed a gradient-boosting decision tree approach using an extensive dataset of 2,159,354 records encompassing complete blood cell counts and biochemical tests, achieving an AUC of 0.998 [26]. However, the study limitations included reliance on simulated mix-up scenarios and single-center validation without external confirmation [26]. Graham et al. focused on a pediatric setting and compared multiple ML models using CBC data from 123,654 patients, with XGBoost again performing best (accuracy of 97.15 %, AUROC of 0.996, sensitivity of 96.90 %, and specificity of 97.39 % for CBC with differentials) [22]. Feature importance analysis revealed that red cell indices consistently ranked among the top 10 predictive features [22].

Comparative analyses of multiple ML architectures have established artificial neural networks as particularly effective for WBIT detection tasks. Farrell and Giannoutsos systematically evaluated eight different ML models using 112,321 samples from a public hospital, including comprehensive CBC parameters, patient demographics, and collection location data [23]. Their artificial neural network achieved superior performance with accuracy of 99.1 %, sensitivity of 99.2 %, and specificity of 98.9 %, consistently outperforming all alternative approaches and exceeding manual review capabilities by laboratory staff. The study’s development of an interpretable simple decision tree using only four CBC parameters – red cell distribution width percentage delta change, mean cell volume percentage delta change, platelet distribution width percentage delta change, and mean cell hemoglobin percentage delta change – provided an implementable solution for existing laboratory information systems [23]. The broader applications of neural networks for specimen identification have been demonstrated across multiple laboratory parameters. Farrell’s investigation of mislabeled sample detection using 127,256 sets of consecutive electrolyte results achieved 92.1 % accuracy with an AUC of 0.977 [24], while a separate study by Farrell [25] utilizing 141,396 sets of electrolyte, urea, and creatinine results demonstrated that artificial neural networks could achieve accuracy of 92.5 %, sensitivity of 90.6 %, and specificity of 94.5 %. Notably, this latter investigation revealed that autonomous AI operation significantly outperformed human-supervised decision support modes, with the autonomous model achieving an AUC of 0.980 compared to reduced performance when human interaction was incorporated. Similarly, deep learning architectures have established new performance benchmarks for specimen mix-up detection through advanced neural network designs. Zhou et al. implemented Deep Belief Network-based delta check methods using 423,290 hematology test results, achieving an AUC of 0.977, accuracy of 93.1 %, true positive rate of 92.9 %, and true negative rate of 93.3 % [28]. Their approach significantly outperformed both reference change value methods and empirical delta checks, though the lack of model explainability was noted as a limitation for clinical implementation [28].

Commercial solutions

Several commercial solutions are available to prevent or detect wrong blood in tube (WBIT) errors and specimen mix-ups during the pre-analytical phase (Table 2). These tools primarily use barcoding, automation, and informatics to ensure that a patient’s sample is correctly labeled and handled. However, none of these uses AI or ML technologies.

Research applications

AI applications in sample dilution management have demonstrated significant potential for optimizing laboratory workflows while reducing resource wastage and improving operational efficiency (Table 1). Ialongo et al. developed a multi-layer perceptron artificial neural network (MLP-ANN) specifically designed to manage sample dilution for serum free light chain testing, addressing the complex decision-making processes required for optimal analytical protocols [29]. Using a comprehensive database of 6,099 entries that included serum free light chain results, dilution status, and patients’ hospital status, their MLP-ANN achieved substantial reductions in wasted tests, decreasing unnecessary serum kappa free light chain testing by 69.4 % and serum lambda free light chain testing by 70.8 % [29].

Commercial solutions

Leading commercial laboratory automation systems from Abbott, Siemens, and Beckman Coulter offer distinct approaches to managing sample dilution, each aiming to improve pre-analytical efficiency and accuracy (Table 2). For example, Abbott’s Alinity system proactively dilutes samples predicted to exceed the analytical range through its middleware, which helps prevent repeat tests and reduces turnaround times. Siemens’ DataLink V2.0 system takes a different approach by incorporating patient-specific clinical data, such as gestational age, to make context-aware dilution decisions. Notably, this system explicitly claims to include AI-assisted intelligent dilution logic that leverages patient data to automate these decisions. However, the system mainly uses rule-based logic rather than a sophisticated AI system [40]. In contrast, Beckman Coulter’s DxA 5000 system employs a comprehensive pre-analytical strategy, prioritizing volume monitoring and quality checks to ensure a sample is adequate before any dilution is attempted.

Research applications

The detection of chemical adulteration in urine samples represents a critical application domain where ML has shown exceptional promise for forensic toxicology and workplace drug testing programs (Table 1). Streun et al. created an ML framework using liquid chromatography coupled to high-resolution mass spectrometry data from 702 authentic human urine samples treated with five different chemical adulterants: pyridinium chlorochromate, potassium nitrite, hydrogen peroxide, iodine, and sodium hypochlorite, with water serving as the control condition [30]. Their artificial neural network achieved accuracy of 90.4 %, sensitivity of 88.9 %, specificity of 92.0 %, positive predictive value of 91.9 %, negative predictive value of 89.2 %, and area under the curve of 0.99. When validated on an independent test set of 202 samples comprising both untreated samples from their study and treated samples from previous research at varying adulterant concentrations, the model achieved 95.4 % accuracy, demonstrating robust performance across different experimental conditions [30].

Commercial solutions

Commercial products designed to detect chemical manipulation in urine samples rely on colorimetric detection technology (Table 2). The underlying principle involves using chemical indicator reactions that cause a visible color change when specific adulterants are present in the specimen. Currently, none of these products explicitly employ AI or ML technologies.

Research applications

The automated assessment of serum quality through AI represents a fundamental advancement in laboratory quality control, addressing the critical need for rapid and accurate detection of hemolysis, icterus, and lipemia interference that can significantly compromise analytical accuracy.

Deep learning architectures have demonstrated exceptional capability for serum quality assessment through advanced image analysis techniques (Table 1). Yang et al. developed a deep learning system using convolutional neural networks based on the Inception-Resnet-V2 architecture to evaluate serum quality from sample images [31]. Their extensive dataset encompassed 16,427 centrifuged blood images with established serum indices values for hemolytic, icteric, and lipemic indices. The model achieved outstanding performance for subclassification tasks, with area under the curve values of 0.989 for hemolysis detection, 0.996 for icterus identification, and 0.993 for lipemia assessment, establishing new benchmarks for automated serum quality evaluation [31]. Similarly, Man et al. developed a deep learning model using ResNet18 architecture for classification of serum samples to detect hemolysis, icterus, and lipemia (HIL) interference, utilizing over 168,000 serum images [32]. The model achieved high reliability with a false negative rate of 2.56 % and significantly improved efficiency with an 86.48 % pass rate (no need for serum index testing) during four months of prospective real-world testing. This approach provides a practical solution for addressing pre-analytical errors in clinical laboratories by reliably identifying HIL samples while maintaining high throughput and reducing reagent costs [32].

Predictive analytics for specific analytical interferences have shown remarkable success in point-of-care testing environments. Benirschke and Gniadek developed a logistic regression model specifically designed to predict falsely elevated point-of-care whole-blood potassium results caused by hemolysis, using data from 3,489 unique patients [33]. Their model achieved an exceptional area under the curve of 0.995, with optimization strategies employing both rule-in and rule-out cutoffs to maximize clinical utility. The rule-in cutoff yielded a positive predictive value of 90.9 % and negative predictive value of 99.0 %, with overall accuracy of 98.9 %, sensitivity of 58.8 %, and specificity of 99.9 %. Conversely, the rule-out cutoff achieved a negative predictive value of 99.9 % and positive predictive value of 61.5 %, with accuracy of 98.5 %, sensitivity of 94.1 %, and specificity of 98.6 %. The model’s biological explainability represented a significant advantage, with elevated creatinine, low carbon dioxide, and increasing ionized calcium consistently associated with true potassium elevations, patterns consistent with renal disease pathophysiology. External validation using 496 point-of-care basic metabolic panels confirmed robust performance, with rule-in cutoff achieving accuracy of 99.0 %, sensitivity of 82.4 %, and specificity of 99.5 % [33].

Commercial solutions

In modern clinical laboratories, automated assessment of hemolysis, icterus, and lipemia (HIL) indices is a key tool for ensuring sample quality before analysis (Table 2). These HIL checks provide a rapid and fully automated way to flag compromised samples before they are tested while none of these well-known systems utilizes AI.

Research applications

ML applications in laboratory test utilization and ordering optimization represent a transformative approach to reducing healthcare costs while improving diagnostic efficiency and clinical decision-making (Table 1).

Specialized test utilization optimization has shown remarkable success in reducing unnecessary procedures while maintaining diagnostic sensitivity. Zhang et al. developed ML models to improve peripheral blood flow cytometry (PBFC) test utilization using data from 784 PBFC samples that included patient history of hematological malignancy and complete blood count differential parameters [34]. Their comparative analysis of decision tree and logistic regression models revealed complementary strengths, with the decision tree achieving 98 % sensitivity and 65 % specificity (AUC=0.906), while the logistic regression model demonstrated 100 % sensitivity and 54 % specificity (AUC=0.919). The decision tree approach provided enhanced explainability through interpretable branching logic. Implementation of these models resulted in substantial reductions in unnecessary PBFC utilization, decreasing inappropriate testing by 35–40 %, though limitations included relatively small sample size and absence of external validation [34].

Large-scale automated laboratory test recommendation systems have demonstrated exceptional performance using deep learning architectures trained on comprehensive healthcare databases. Islam et al. developed deep neural network system using retrospective data from the Taiwanese National Health Insurance Research and Development database, encompassing 1,463,837 prescriptions from 530,050 unique patients [35]. Their three-layer deep neural network achieved outstanding performance with area under the receiver operating characteristic curve (AUROC) of 0.98, while individual laboratory tests demonstrated AUROC values ranging from 0.76 to 1. However, the deep learning approach was characterized by limited interpretability, often described as a “black box” system, with additional limitations including absence of external validation [35]. Another study from Islam et al. focused on cardiology department patients [36]. They developed deep learning models trained on 35 laboratory tests that represented more than 90 % of total tests ordered in this clinical setting. Their deep neural network achieved an AUROC of 0.87, with 99 % sensitivity and AUROC ranges from 0.63 to 0.90 across different laboratory tests. While this specialized approach enabled more targeted clinical application, limitations included restriction to a single clinical department and relatively limited test menu coverage, highlighting the challenges of developing generalizable algorithms across diverse clinical contexts [36].

Intelligent reflex testing represents an innovative application of ML to automate complex clinical decision-making traditionally requiring expert knowledge. McDermott et al. developed “smart” reflex testing protocols using ferritin testing as a representative example, analyzing 288,427 unique complete blood count events across 137,451 unique patients [37]. Their random forest model achieved an AUC of 0.731. Feature importance analysis revealed clinically relevant predictors for iron deficiency, with previous ferritin testing within two years serving as a highly informative predictor. Model predictions aligned more closely with ideal clinical practice than actual physician ordering patterns, though limitations included subjectivity of clinical validation and single-center study design [37].

Commercial solutions

Commercial tools have emerged to help laboratories optimize test utilization and improve ordering practices (Table 2). These solutions leverage data analytics, decision support, and automation to reduce unnecessary tests, ensure appropriate ordering, and enhance cost efficiency. While many of these products use sophisticated algorithms to analyze testing patterns and generate insights, the hc1 platform is the most explicit in its use of AI. It employs ML to proactively flag and address test misutilization.

Research applications

AI applications in automated tube handling and labeling represent sophisticated integrations of computer vision, robotics, and ML technologies designed to enhance preanalytical efficiency while reducing human error rates (Table 1).

Demirci evaluated the operational performance of NESLI, an AI-based tube-labeling robot incorporating two convolutional neural networks [38]. The first neural network processed overhead camera images to precisely determine tube positions, while the second network classified tubes based on cap color characteristics. Using simulator-generated data based on 12 months of real-world operational patterns from hospital blood collection units, incorporating 330 synthetic orders containing varying combinations of standard-sized and nonstandard-sized tubes, the system achieved 99.2 % success rate for labeling parameters related to manufacturer maximum blood level mark visibility and 100 % success across seven additional labeling quality parameters, including consistency between selected tubes and orders, quality of label application, consistency in external label printing, consistency between label and barcode data with orders, readability of printed barcodes, and recovery of barcode data by standard scanners. Other operational parameters achieved 100 % success rates for tube handling, critical stock warnings, and physical tube location and gripping functions, with median labeling time of 8.96 s per tube. System limitations included tube storage capacity restricted to 204 tubes and occasional temporary pauses due to hospital network interruptions [38]. Similarly, Şişman et al. evaluated KANKA, an AI-driven robot designed for blood tube sorting and preanalytical quality control, utilizing three specialized convolutional deep learning models complemented by a barcode scanning system [39]. The first model determined exact tube positioning, the second identified tube type and assessed adequate blood volume, and the third model specifically identified filling levels, while the fourth model for barcode scanning functionality. Using an experimental validation approach with 1,000 blood tubes from different patients loaded five times to generate 5,000 total trials, KANKA achieved an accuracy of 99.98 % in quality control and sorting operations. The system demonstrated 100 % success rate in detecting all categories of preanalytical errors, including tubes without barcode labels, incorrect barcodes, wrong tube types, empty tubes, tubes intended for other laboratories, tubes without appropriate colored caps, and tubes with inadequate blood volumes. Processing efficiency reached 311 blood tubes per hour, with the capability to store and retrieve sample images providing additional confirmation of AI results, though limitations included the preliminary experimental nature of the validation, exclusion of hemolysis detection, and absence of pediatric tube testing in the current study protocol [39].

Commercial solutions

The aforementioned AI-powered tube handling systems automated tasks that laboratory technicians perform manually. However, several automated pre-analytical systems have been developed to enhance tube handling and labeling, ensuring that samples are correctly collected, labeled, and processed with minimal errors without requiring AI or ML algorithms (Table 2).