Hilab system, a new point-of-care hematology analyzer supported by the Internet of Things and Artificial Intelligence
Sample preparation process
The Hilab device accepts both venous and fingerstick blood samples. Two drops of blood, totaling 90 l, are necessary for test realization. The sample is collected directly from the finger (or K3EDTA tube to venous workflow), using the components provided in the test kits. Two single-use diagnostic kits are utilized for the realization of point-of-care CBC tests. The first, used for cell counting, presents a disposable hemocytometer, the diluent solutions, blood collection pipettes, and blood transference pipettes. The second, used for hematimetric parameters evaluation, contains a chromatographic strip and a blood collection pipette (Fig. 1B,E, respectively). Both diagnostic kits present the materials needed for the capillary puncture.
Blood cell count
Two diluent solutions are needed to dye the cells and dilute the sample for blood cell counting. During the sample preparation, the test operator collects a drop of blood (40 uL) and places it into solution 1 (9-fold blood dilution). Solution 2 goes through the same process (179-fold blood dilution). After the blood homogenization, the test operator individually transfers a drop (10 l) of solutions 1 and 2 to the hemocytometer chambers (Fig. 1C). The hemocytometer insertion into the Hilab Lens device is the last operation step.
The composition of the first diluent solution includes dyes, salts, and surfactants that promote RBC lysis and WBC differential stain. On the other hand, the second one is composed of different salts, which keep the natural morphology of RBC and PLT. Both solutions were developed and patented by Hilab and allow the liquid medium blood cells observation by optical microscopy.
Hematometric parameters evaluation
For hematimetric parameters evaluation, a single drop of blood (10 l) is collected and deposited on the chromatographic strip contained in a plastic capsule. This strip presents a fixed reagent that promotes RBC lysis and HB conversion into methemoglobin. The process is concluded with the insertion of the capsule into the Hilab Flow device for HB quantification (Fig. 1F). In this step, the test operator does not realize blood dilution. From obtained HB and RBC values, the artificial intelligence (AI) estimates the HT, MCV, and MCH, based on previous studies (FAILACE and FERNANDES, 2015). The principle of these calculations is the correlation between HB and/or RBC values, with HT, MCV, and MCH values, allowing the estimation of these analytes.
The Hilab system uses microscopy and chromatography techniques to supply the CBC result. The microscopy one is handled by a small handheld device (19.7 × 9.9 × 15.3 cm; 0.5 kg) called Hilab Lens (National Health Surveillance Agency (ANVISA) registration no. 80583710018; Fig. 1A). The acquisition process occurs by the autofocus and image capture process, which takes upwards of 400 images of each sample to form the final image by the composition of all figures stacked. This device processes the blood sample in two stages: the first is used to read the WBC (first chamber), and the second is to read RBC and PLT (second chamber). Considering the standard dilution factor used in cell counting single-use kits, the number of cells that this device analyzes in each sample is dependent on the blood cell number of the patient. The chromatography technique is handled by a small handheld analyzer (12.4 × 12.4 × 12.7 cm; 0.45 kg) called Hilab Flow (ANVISA registration no. 80583710007; Fig. 1D). This device incorporates a camera-equipped light detector and sample integrated capsules that enable the processing of several analytes by the optical density of chromatography strips.
Through the calibration capsule, the test operator calibrates both devices every 24 h to verify the correct functioning of the sensors and the position of the focus mechanism. Hilab Flow and Lens need energy access for battery charging and an internet connection to enable the Hilab patient registration platform access. The patient registration data are encrypted and linked to the device and the exam QR code through a key. After sample processing, the Hilab System sends the patient data, test realization device, and the test operator register to the Hilab Software discussed below. Hilab uses two non-relational databases to protect the patient data and test results, precluding the information correlation in case of a system invasion.
Imaging processing and object classification
For Hilab Lens, blood cell images go through a deep learning approach for both cell detection and classification. We applied data augmentation to RBC, PLT, and WBC subpopulations, generating more data through rotating and mirroring existing images. For cell detection, given that a single image covers thousands of cells—small objects—the image is divided into a grid and each sub-image is processed individually through a pipeline with the YOLO (You Only Look Once) deep learning approach, a state -of-the-art real-time object detection system that analyzes the image globally only once, using features from the entire image, to predict each bounding box (a box containing positions delimiting the detected object). The algorithm used for this analysis includes an overlap verification, in which detections that have less than a threshold of intersection are considered the same cell. Since YOLO also predicts class probabilities for each detected object, the network is also used for pipeline classification through the probability evaluation that a given region contains a specific type of object.
For hematimetric parameters, the Hilab Flow uses AI and computer vision methods for HB detection on the chromatographic strip. Thus, after the device collects the signals of the blood sample, signal processing techniques extract the sample mean colorimetric value, and the system applies regression analysis to obtain the HB concentration. Finally, from RBC and HB results, the Hilab System estimates the HCT, MCV, and MCH.
Processing of results
After AI image processing, Hilab’s software receives the figures through the internet connection. The software is an interactive platform developed by Hilab to connect the habilitated health professional to the AI processed image of the blood sample. Only our trained professionals, with active registration in the professional class council, have access to this platform in which the professional double-checks the AI results of blood samples and makes corrections, if necessary. If any divergence occurs during the exam analysis, the result provided by the habilitated professional is always prevalent about AI
The number of blood cells is obtained through the cell count considering the double-checked processed image, the standard dilution factor used in cell counting single-use kits (179-fold blood dilution for erythrocytes and platelets; 9-fold blood dilution for leukocytes) , and the standard volume of the high precision hemocytometer chamber. Finally, the health professional issues the report, and the patient receives it signed by Hilab technical responsible and the analyst of the report, through email or SMS. Thus, the AI accuracy guarantees fast analysis, and the specialists double-check results ensures a reliable exam report. The sample preparation, imaging acquisition, and processing of results by the Hilab system take an average time of 30 min. Through the Hilab software the health professional has access to the exam lot, test realization device, test operator register, and patient exam history. Hilab guarantees 24-7 service to the correct system function, reports delivery, and customer support.
The Research Ethics Committee of the Paranaense League Against Cancer (CAAE; no. 49961421.3.1001.0098) approved this study, and all performed methods were in accordance with the relevant guidelines and regulations. Venous whole blood clinical samples (N = 450) were collected from patients aged between 0.6 and 86 years old, including males (42%) and females (58%), by trained and qualified professionals. All subjects or their legal guardian(s) informed consent for study participation. The samples encompassed normal (82%) and pathological conditions (18%), such as thalassemias (2.2%), anemias (6.6%), and infections (9.2%). The venous whole blood samples were stored in standard K3EDTA collection tubes (Vacuette, Greiner Bio-One, Brazil) and processed within 12 h of collection. The sample processing included blood analysis in the Hilab system and the standardized Sysmex XE-2100 analyzer (Sysmex Corporation, Japan; reference values). Although traditional flow cytometry is considered the gold standard method for population cell differentiation, mainly due to the cost, most clinical laboratories use devices with the impedance-resistivity methodology for CBC realization, justifying the analyzer choice. The analyzes of the Hilab System and Sysmex XE-2100 analyzer results were done by different professionals, being a double-blinded study.
Pearson correlation, Student t test, bias, and the Bland–Altman plot of each blood count analyte were calculated and shown. Also, we compared the Hilab System accuracy of each CBC analyte to the Sysmex XE-2100. Thus, we assessed each value inside (1) or outside (0) of the reference range through the confusion matrix. The evaluated parameters were specificity, sensitivity, kappa coefficient, and balanced accuracy. All biological samples collected were single-use for this study and discarded after the analysis, following the potentially infected samples standard procedure.
K3EDTA whole blood venous samples were used for precision study due to the incompatibility of commercial hematological controls with the Hilab solutions. To encompass the different clinical ranges of CBC analytes, four extended ranges of each PLT, WBC, RBC, and HB sample, were measured at consecutive times in three devices. For the repeatability study, within-day precision was evaluated by each range standard deviation (SD) and within coefficient of variation (CV). For the reproducibility study, this protocol was performed by two different operators and evaluated for three consecutive days. Similarly, the SD and CV evaluations of each clinical range were done.
Equivalence between capillary and venous samples
Fresh fingerstick blood samples were collected from healthy volunteers (n = 150) parallelly with venous blood collection. For all cell blood counts, results from capillary samples were compared with the respective venous plus anticoagulant (K3EDTA) samples using the Passing–Bablok analysis and paired Student t test.
The flagging capabilities of the Hilab system were compared to the manual microscopy technique, sending the blood smears to the analysis of trained personnel from a support laboratory (Diagnostico do Brasil, Parana, Brazil). The evaluation of RBC morphological abnormalities, including microcytosis, anisocytosis, and macrocytosis, was realized. We also evaluated the PLT and WBC abnormalities, assessing the presence of platelet clumps and immature cells, respectively. The accuracy, specificity, sensibility, kappa coefficient, and balanced accuracy of each parameter were calculated and expressed.
All data were analyzed and plotted using the R software statistics package analysis. The Kolmogorov–Smirnov normality test was applied to ensure that the data met the criteria for performing the parametric tests. CV, SD, Bias, Student t test, Bland–Altman, and Passing–Bablok analysis were calculated using this package. The significance level was set at p 0.05.
Ethics approval and consent to participate
This study was approved by the Research Ethics Committee of the Paranaense League Against Cancer (CAAE no. 49961421.3.1001.0098).