An artificial intelligence‐assisted diagnostic platform for rapid near‐patient hematology

Abstract Hematology analyzers capable of performing complete blood count (CBC) have lagged in their prevalence at the point‐of‐care. Sight OLO (Sight Diagnostics, Israel) is a novel hematological platform which provides a 19‐parameter, five‐part differential CBC, and is designed to address the limitations in current point‐of‐care hematology analyzers using recent advances in artificial intelligence (AI) and computer vision. Accuracy, repeatability, and flagging capabilities of OLO were compared with the Sysmex XN‐Series System (Sysmex, Japan). Matrix studies compared performance using venous, capillary and direct‐from‐fingerprick blood samples. Regression analysis shows strong concordance between OLO and the Sysmex XN, demonstrating that OLO performs with high accuracy for all CBC parameters. High repeatability and reproducibility were demonstrated for most of the testing parameters. The analytical performance of the OLO hematology analyzer was validated in a multicenter clinical laboratory setting, demonstrating its accuracy and comparability to clinical laboratory‐based hematology analyzers. Furthermore, the study demonstrated the validity of CBC analysis of samples collected directly from fingerpricks.


| INTRODUCTION
A large number of laboratory tests are now available for point-ofcare testing (POCT) 1 including, among others, urinalysis, blood chemistry and infectious disease testing. [2][3][4][5] However, hematology analyzers capable of performing complete blood counts (CBCs) have lagged in their prevalence at the point-of-care, 6 despite CBCs ranking among the most commonly ordered tests. The relative absence of POC CBC analyzers has also hampered the adoption of other POC tests that tend to be ordered alongside the CBC (e.g., complete metabolic panels); in particular, if a CBC is required but must be performed in the central lab due to its absence at the point-of-care, its companion tests may as well be ordered from the central lab.
The majority of current near-patient CBC analyzers rely on the miniaturization of large-lab analyzers, particularly those based on flow cytometry and/or impedance cytometry techniques. 1,4,7 Such miniaturization has in many cases been accompanied by simplification, which has resulted in reduced numbers of parameters (e.g., three-part instead of five-part differentials), narrower reportable ranges, reduced abnormal cell flagging capabilities and in some cases a reduction in accuracy (as indicated by the small number of FDA clearances). [8][9][10] These CBC analyzers have also inherited many of the product attributes of their larger relatives, including the requirements for liquid reagent replacement, washout and calibration procedures, and frequent quality-control processes 11 which in the POC context can be limiting or prohibitive.
Sight OLO (S.D. Sight Diagnostics LTD, Israel) is a novel hematological platform designed to address some of the limitations in current near-patient and POCT hematology analyzers. Note, OLO is based on recent advances in artificial intelligence (AI) and computerized image analysis (computer vision), and it provides a 19-parameter, five-part differential CBC. And, OLO employs single-use test-kits, a "dry" instrument and an extensive "Failsafe" self-test system to reduce operation overheads and simplify operation. Sight OLO has received FDA 510(k) clearance for use in CLIA non-waived settings 12 and is CE marked for POC use, including its use with samples collected directly from fingerpricks. Here, the results are presented for the accuracy of OLO as compared with the Sysmex XN-Series System (Sysmex, Kobe, Japan), as well as repeatability, flagging capabilities, reproducibility and matrix studies comparing the venous, capillary and fingerprick blood samples.

| The device
Sight OLO is a desktop hematology platform that was designed from the ground up for POC and near-patient settings. The device employs single-use test kits that use a novel method for creating and staining blood smears within disposable test cartridges (please see "sample preparation" below). These monolayer blood smears are rapidly imaged using OLO's automated fluorescence microscope to yield a set of more than 1000 multispectral micrographs per sample.
The images are analyzed using OLO's onboard computer as they are collected using a set of specially designed algorithms, with the analysis resulting in the 19 reported CBC parameters and several flags, including flagging WBC for the presence of nucleated RBCs, blast cells, immature granulocytes and atypical lymphocytes. Then OLO displays the diagnostic result on its touchscreen interface, optionally provides a paper printout and transmits the result digitally to a laboratory information management system (LIMS) or electronic medical records (EMR) system. The device measures 32 cm by 28 cm by 25 cm (about one cubic foot) and houses no reagents; reagents are contained within each test kit, obviating the need for regular instrument washouts. Sight OLO is factory calibrated so it requires no regular calibration. By ensuring that all the optical, mechanical and electronic elements of the system that may have a direct effect on the analyzer calibration (for example: LED intensities, illumination maps, focus mechanism, cartridge leveling etc.) remain within allowed ranges and enforced by a lock-out method, Sight OLO's Failsafe system works as an internal quality control to ensure that the device is performing appropriately, thereby obviating external quality control (QC) materials except as mandated by regulation.
In addition, the Failsafe system is designed to mitigate user errors, consumable defects and blood sample irregularities by rejecting such samples to ensure that the analyzer does not return erroneous results.

| Sample requirements and preparation process
Sight OLO accepts both venous and fingerprick blood samples. In the fingerprick workflow, two drops of blood totaling 27 μl are collected directly from the finger, using the components provided in each single-use test kit as described below. Sight OLO was designed to work with low sample volumes to permit sample collection with no "milking" of the finger, in order to minimize patient discomfort. In the venous workflow (which also supports capillary samples collected in low-volume collection tubes), blood is provided in a K 2 EDTA tube; the sample is deposited from the collection tube onto hydrophobic paper using standard tube-top dispensers (e.g., Labcon U-Pette) and then prepared using the components of the same test kit and following the same process as fingerprick samples.
Each Sight OLO single-use test kit contains a test cartridge, two capillaries for sample collection and a reagent-filled mixing-bottle. In turn, the test cartridge contains two separate sample chambers: a hemoglobin chamber and an imaging chamber. During sample preparation, a single drop of undiluted blood (17 μl) is used to fill the hemoglobin chamber; in the fingerprick workflow, this drop can be collected directly from the pricked finger using the provided capillary.
An additional 10 μl of blood is collected using the second capillary (K 2 EDTA coated) and mixed with a diluent and dried fluorescent stains within the mixing-bottle. The resulting blood mixture is then used to fill the imaging chamber ( Figure 1).

| Live monolayer imaging
To facilitate high-resolution microscopy of individual cells, the blood sample has to be presented for imaging as a monolayer -a film characterized by a defined focal plane in which cells rarely overlap with each other. Traditionally, this is accomplished through the preparation of blood smears; however, since the quality of blood films prepared using this process tends to vary significantly between users and across the sample area, the method was deemed unsuitable for POC usage.
Instead, Sight OLO relies on a novel monolayer formation process: the diluted blood sample is drawn using capillary action into the imaging chamber, which has a predefined height of a few hundred of microns. Since blood cells are denser than the diluent, they gradually sink and settle on the chamber floor. Due to the small height of the chamber, this process only takes about 1 min, leaving the cells at a defined focal plane as desired for imaging. Furthermore, the dilution ratio is selected to ensure that, for the most part, the settled cells do not overlap regardless of the concentration of red blood cells in the sample (which ranges roughly between 2 Â 10 6 /μl and 8 Â 10 6 /μl).
No sphering or fixation reagents are included in the diluent, in order to retain cell morphology.

| Staining and multispectral imaging
Traditionally, blood smears are stained with Wright-Giemsa stains and their variants, which have a century-long track record in the differential identification of peripheral blood cell populations. However, these stains suffer from a number of drawbacks, including significant incubation times, the requirements for washing steps, the need for fixation (which disrupts morphology) and variability in stain

| Scanning hardware
Sight OLO contains a fully automated fluorescence microscope which rapidly collects high-quality images of the blood smear present within its single-use test cartridge. The analyzer automatically scans at least 200 non-overlapping fields within the sample, and at each field it acquires the several channels of fluorescent and brightfield imaging. In order to simplify manufacturing constraints on cartridge flatness, OLO automatically refocuses on the sample at every field. Each blood sample is thus "digitized" into approximately 6 GB of image data.
In addition to microscopy, the Sight OLO also measures hemoglobin using an unlysed, reagent-free process utilizing four wavelengths to account for absorption and scattering. 13 The hemoglobin chamber is directly filled with undiluted whole blood using capillary action and contains several measurement areas, which differ by optical path length. These areas are used to derive differential measurements in order to normalize for system parameters such as illumination intensity and manufacturing tolerances.

| Platelets
Platelets are detected using fluorescent staining. However, due to their low RNA content, platelets require a longer exposure time in order to obtain a strong enough signal to stand above the background.
This signal is combined with the brightfield channels to detect the candidates. However, being much smaller and less bright than the other two cell types, cell fragments and debris are sometimes also detected as platelet candidates. Accordingly, true platelets are identified first by filtering the candidates according to different morphological and intensity properties, then applying several convolutional neural networks trained to accurately distinguish the platelets from background in different scenarios.  16 and is consistent with the approach conducted for FDA submissions for hematology devices.

| Repeatability study
Within-run repeatability studies were performed using residual K 2 EDTA whole blood venous samples (minimum of 2 ml per sample). In order to span healthy and pathological values and the medical decision points between them, each site tested at least four samples within lab reference ranges, three samples around medical decision levels for HGB (6-10 g/dl), PLT (<50 Â 10 3 /μl) and WBC (<2 Â 10 3 /μl), and four samples around the upper range for RBC (>6 Â 10 6 /μl), HGB (>17 g/dl), WBC (>12 Â 10 3 /μl) and PLT (>600 Â 10 3 /μl). In total, this requirement led to 38 samples being scanned, with each sample measured 20 consecutive times (after excluding invalidations or rejects). Standard deviation (SD) and coefficient of variation were calculated for each run. The first 20 successful runs per measurand and were analyzed. If fewer than 20 successful scans were obtained within the required time slot from phlebotomy (8 h), the sample was still analyzed so long as 17 or more replicates were scanned. For the anemic samples (HGB 6-10 g/dl) only RBC, HGB and HCT were analyzed, while for the thrombocytopenic (<50 Â 10 3 /μl) and leukopenic (<2 Â 10 3 /μl) samples, only PLT and only the WBC concentration and differential were analyzed, respectively. This was in accordance with the CLSI guidance H26-A2, 15 which refers to the ICSH protocol for evaluation of blood cell counters.

| Reproducibility studies
Reproducibility studies were conducted using three levels of commercial control materials (low, normal and high -below, within and above the reference ranges of main parameters, respectively) for all reported

| Flagging study
The flagging capabilities of Sight OLO were compared to manual microscopy for WBC distributional abnormalities and WBC morphological abnormalities, which include blasts, immature granulocytes, nucleated RBCs, and atypical lymphocytes. Other invalidating messages, such as platelet clumps and RBC agglutination were not included in the flagging study. The 108 negative and 100 positive samples with WBC count larger than 4 Â 10 3 /μl were enrolled from the method comparison study. For each sample, the three prepared smears were sent to analysis by trained morphologists from Columbia University Irving Medical Center staff, who had no access to either clinical information or reference method results. The testing design was based on the test methods outlined in CLSI H20-A2 14 and CLSI H26-A2. 15

| Matrix comparison study
A matrix comparison study was performed in two parts to assess the equivalence between venous and capillary samples, and between capillary and direct-from-finger samples. In the first part of the study, 67 samples (of 52 subjects, 15 of which repeated the test after a 6-month interval) were collected in pairs. A venous sample was drawn, and a capillary sample was collected into 350 μl microtainers. Healthy volunteers were primarily tested, and additional samples at medical decision points and across the analytical measuring range (i.e.,: HGB 6-10 g/dl; WBC <2 Â 10 3 /μl; PLT <50 Â 10 3 /μl; WBC >12 Â 10 3 /μl; PLTs >500 Â 10 3 /μl; RBCs >6 Â 10 6 /μl; HGB >17 g/dl) were enrolled from subjects in Tel Aviv Sourasky Medical Center.

| Method comparison
The accuracy of Sight OLO was compared with the Sysmex XN-1000 System. The study design was based on the methods outlined below and in accordance with CLSI H20-A2, 14 CLSI H26-A2 15 and CLSI EP09-A3. 16 Samples from patients age 3 months to 94 years and included 355 males (52%) and 324 females (48%) were analyzed; 32% of the samples were from pediatric patients (3 months-21 years).
Samples were selected to comprise several abnormalities, including different blood disorders and tumors (e.g., various types of anemias, leukemias, lymphomas, myelomas), and covered a wide clinical range for each of the tested parameters.
The results of the regression analysis, which are included in Figure 2, show a strong concordance between the Sight OLO and the Sysmex XN both in terms of correlation coefficient and as seen through slope, bias and intercept. Sight OLO performs with high accuracy for all CBC parameters. Detailed results are included in Table S1. Figure S1 and Table S2 illustrate a version of the analysis without the exclusion of measurands invalidated by OLO, demonstrating the high number of actionable results.

| Repeatability and reproducibility
Within-run repeatability studies were performed using residual K 2 EDTA whole blood samples as described below. In these studies, Sight OLO demonstrated high repeatability for most of the testing parameters. Table 1 shows the pooled SD for each measurand in relevant clinical ranges across all samples within that range. Reproducibility results are included in the Table S4 and show similar performance to the within-run whole blood repeatability shown in Table 1.  Figure 2). In the differential, a very high correlation (r > 0.98) was seen for neutrophils, lymphocytes and eosinophils whereas a moderately high correlation (r = 0.89) was found for the monocytes fraction and a r = 0.67 correlation was found for the basophils faction. Also, low correlation was found for mean corpuscular hemoglobin concentration (r = 0.69). The slightly low correlation for basophils is not uncommon, as the typical data range for this parameter is very limited, leading to low correlation values. The slightly low correlation for MCHC is due to the fact that in both OLO and the Sysmex XN, MCHC is a calculated parameter based on the ratio of two highly cor- Results indicate that for these flags the OLO system was in high accordance with manual microscopy.

| White blood cell flagging study
Sight OLO offers a full CBC utilizing only 27 μl of blood sample, providing substantial blood conservation for oncology, neonatal and pediatric patients. 19 While OLO is currently indicated for use in patients 3 months and older, further studies are currently being conducted to increase the population range. Furthermore, the fact that 27 μl may be directly drawn from a fingerprick reduces some of the additional overheads and inconveniences associated with venous phlebotomy or larger volume capillary collection and may alleviate the anxiety involved with these for certain apprehensive populations such as pediatrics or patients who poorly tolerate traditional needle-based blood drawings. The disposable cartridge utilized by OLO has added benefits, as the risk of carryover or system clogging are eliminated, and no maintenance and cleaning are required between runs.
The study presented here successfully validated the performance of Sight OLO hematology analyzer in a multicenter clinical laboratory setting. In particular, the study demonstrated that OLO is accurate and comparable to the renowned Sysmex XN Series Hematology Analyzers. This study led to OLO's FDA 510(k) clearance, 12 and it demonstrates the capabilities of multi-spectral live monolayer imaging and AI-assisted image analysis in hematology.
The study also demonstrated the validity of performing five-part differential CBC analysis using direct-from-finger blood samples: Sight OLO was able to produce results from fingerpricks that are equivalent to those obtained using venous blood draws. Of note, OLO is the first CBC analyzer to be cleared by the FDA for collecting samples directly from fingerpicks.

ACKNOWLEDGMENTS
Research funding: S.D. Sight Diagnostics LTD.