A microfluidic hemostatic diagnostics platform: Harnessing coagulation-induced adaptive-bubble behavioral perception

Summary

Clinical viscoelastic hemostatic assays, which have been used for decades, rely on measuring biomechanical responses to physical stimuli but face challenges related to high device and test cost, limited portability, and limited scalability.. Here, we report a differential pattern using self-induced adaptive-bubble behavioral perception to refresh it. The adaptive behaviors of bubble deformation during coagulation precisely describe the transformation of viscoelastic hemostatic properties, being free of the precise and complex physical devices. And the integrated bubble array chip allows microassays and enables multi-bubble tests with good reproducibility. Recognition of the developed bubble behaviors empowers automated and user-friendly diagnosis. In a prospective clinical study (clinical model development [n = 273]; clinical assay [n = 44]), we show that the diagnostic accuracies were 99.1% for key viscoelastic hemostatic assay indicators (reaction time [R], kinetics time [K], alpha angle [Angle], maximum amplitude [MA], lysis at 30 min [LY30]; n = 220) and 100% (n = 44) for hypercoagulation, healthy, and hypocoagulation diagnoses. This should provide fresh insight into existing paradigms and help more clinical needs.

Graphical abstract

Highlights

• •Explore and reveal the adaptive-bubble behavior in dynamical coagulation processes
• •Develop a microbubble biosensing platform allowing adaptive behavior acquisition
• •Harness bubble behavioral recognition to enable coagulation monitoring and diagnosis

Chen et al. report a method for hemostatic diagnostics employing self-induced adaptive-bubble behavioral perception. The study discovers and reveals the microbubbles’ adaptive behavior in dynamical coagulation processes, harnessing this behavioral recognition to enable coagulation function diagnosis. It should provide fresh insight into existing paradigms and help more clinical needs.

Introduction

The viscoelastic hemostatic assay, the description of biomechanical properties’ transformation during coagulation, has crucial significance for much clinical decision-making, such as trauma-induced coagulopathy, resuscitation guide during transfusions, etc.^1,2,3,4 Because it enables monitoring of the entire coagulation process from the beginning of clotting to clotting formation and fibrinolysis, it provides a complete assessment of coagulation factors, fibrinogen, platelet aggregation, and fibrinolysis.^5,6 For example, in trauma resuscitation, the early availability of the viscoelastic hemostatic assay’s data will be important to guide hemostatic decisions, which could improve case survival,^7,8 and it is urgently desired in public health events and emergency situations (emergency rescue, disaster relief, battlefields, etc.).

Over decades of development, several excellent clinical instruments have been proposed for clinical viscoelastic hemostatic assays, including thromboelastography (Haemonetics, Boston, MA, USA) and thromboelastometry (Instrumentation Laboratory, Munich, Germany),^4,5,6,7 but the current bottleneck is that the available clinical methods are dominated by physical stimulation-response systems (probes, resonance modules, etc.), leading to their high device/test cost and limited portability and caseload capability. New device/method and breakthrough technology development is eagerly awaited to empower convenient assays in diverse clinical settings. Currently, natural bubbles are of increasing interest due to their extensive presence and impact from industrial to living systems, their excellent biocompatibility, and their size span, which ranges from the macro- to the micro-nanoscale, endowing its ability for on-demand tunable biological applications.^8,9 And their adaptive behaviors (stretching, collapsing, etc.) are often accompanied by features typically associated with mechanic/elastic changes,^8,9,11 hinting at their tremendous potential for biomechanical sensing. Besides, emerging studies show that bubble systems could integrate with microfluidics and that they have contributed to the study of dynamical fluid mechanics sensing, tissue engineering, drug screening, etc.,^8,9,11,12,13 which also provides a potential platform for dynamical monitoring of the coagulation process.

Here, we report a method using self-induced adaptive-bubble behavioral perception (Figures 1 and 2A) for hemostatic diagnostic. The viscoelastic hemostatic properties’ transformation is bridged with adaptive-bubble behaviors to enable passive assays without active physical source need. The integrated microbubble array chip allows easy operation and microassays with smaller blood samples.^14,15,16 A set of images of dynamical bubble behaviors are then analyzed on a portable edge-computing module^{17,18,19,20,21,22,23,24} via custom-developed algorithms^18,20,24 for automated and user-friendly diagnosis.²⁵ We then prospectively tested the method in human subjects for usability and accuracy against a standard clinical device. Our results support clinical equivalency in accuracy.

Figure 1.

The self-induced adaptive-bubble behaviors during blood coagulation

(A) The green fluorescence refers to fibrin, red fluorescence refers to platelets, and white arrows refer to the direction of bubble deformation behaviors. Scale bar: 50 μm.

(B) Imaging analysis of fibrin and platelets fluorescence. Scale bar: 50 μm.

(C) The regional fluorescence analysis around bubbles.

(D) The fluorescence intensity distribution around deformed bubbles. Error bars represent standard deviation.

(E) The minor and max axes of the deformed bubbles. Error bars represent standard deviation.

Figure 2.

The schematic diagram of working principle and technology details, and a photo of the developed analyzer

(A) The schematic diagram of clotting process at the air-blood interface in epidermal trauma.

(B) Photo of the developed analyzer for viscoelastic hemostatic assays.

(C) The schematic diagram of on-chip adaptive-bubble behaviors during blood coagulation.

(D) The image analysis for bubble behavior properties acquisition.

(E) The acquired bubble behavior properties for clinical indicators’ diagnoses.

Results

The system design

The system (Figures 2B and S1) is equipped with (1) the microfluidic bubble array chip for viscoelastic hemostatic assays (Video S1. Perfusion process, related to Figure 4, Video S2. Bubble behaviors in anti- and healthy coagulation, related to STAR Methods, Video S3. Bubble behaviors without a chip seal, related to STAR Methods), and a coverslip was put on the surface of the chip for oil sealing^16,26,27,28; (2) a developed imaging system¹⁸ for adaptive-bubble behavior imaging; (3) a portable and miniaturized temperature control system including a heating film (12 V, 3 W, 48 Ω, Wuxi Tianbo Electrical Appliance Manufacturing, Jiangsu, China), a temperature controller (XY-WT03-W, OUCHEN, Guangzhou, China; accuracy: 0.1°C); (4) a Raspberry Pi 4 Computer Model B module with custom-developed algorithms for automated and user-friendly diagnoses and (5) a wireless module for pathology results sharing; and (6) a touchscreen for user-friendly operation (Video S4). The device is powered by a lithium ion battery pack for portable operation (Figure S2). The overall dimensions are 160 (L) × 110 (W) × 122 mm (H), and it is 1.1 kg in weight. The microfluidic bubble array chip has been shown in Figure 2C, and the per microfluidic chip has 3 bubble columns (300 [L] × 70 [W] × 100 μm [H]). The developed imaging platform captures images of bubble behaviors during coagulation, and the custom-developed algorithm (Figure 2D) is used for calculating and acquiring the bubble behavior properties (stretching area of microbubble column).^16,18 The bubble behavior properties are used for curve plotting and key clinical viscoelastic hemostatic indicator (reaction time [R], kinetics time [K], alpha angle [Angle], maximum amplitude [MA], lysis at 30 min [LY30]) diagnoses. The acquired clinical indicators are then loaded as input vectors into the developed neural network for hypercoagulation, healthy, and hypocoagulation diagnoses (Figure 2E).

On-chip adaptive-bubble behaviors during blood coagulation

As shown in Figure 1, blood containing random bubbles was injected into a microfluidic chamber (1,000 [L] × 1,000 [W] × 100 μm [H]). The chip was then sealed,^16,26,27,28 and the adaptive deformations of the bubbles and fluorescence images were recorded and analyzed. The results and imaging analysis demonstrate that the bubbles would undergo adaptive deformation (changing from a circular shape to an elliptical shape). By observing the coagulation process, we found that the clotting process would produce negative pressure due to the contraction of blood cell volume during coagulation^29,30,31 and would induce adaptive-bubble behaviors (Figures S8 and S9). To avoid the common evaporation issue in open systems,³² in the case, this system has been sealed,^16,26,27,28 and the temperature remains constant. The bubble pressure ($P_{bub}$) can be expressed by the Young-Laplace equation^8,9 as

$$P_{bub}=P_O+P_B+P_{YL}\text{,}$$

where $P_O$ is the atmospheric pressure, $P_B$ is the blood-coagulation-produced negative pressure, $P_{YL}$ is the Young-Laplace pressure, and $P_{YL}=K\gamma$, where $\gamma$ is the liquid’s surface tension and $K$ is the bubble curvature (K = 1/R, R refers to the curvature radius). The initial state can be described as $P_{bub}-P_O-K\gamma =0$. As the clotting began, the clot growth would produce negative pressure (P_B), and the formula would be changed to $\gamma /R$ = $P_{bub}-P_O-P_B$. From the above equation, we can reason that $\gamma /R$ is increasing, which indicates a decreased curvature radius. Specially, the bubble would be stretched.

The fluorescent characterization was then performed to display clearly and dynamically the on-chip adaptive-bubble behaviors during coagulation. As the clotting process proceeds, blood cells transition from a dispersed state to an aggregated state, and fibrin is generated and entangled. This progress would cause an increase in both the fluorescent area and the intensity.^16,33 The continuous confocal micrographs of on-chip adaptive-bubble behaviors are shown in Figure 3A, and the three-dimensional fluorescence intensity analysis is shown in Figure 3B. The platelet-fibrin co-localization (Pearson’s coefficient) during blood coagulation is shown in Figure 3C.^33,34,35 These results show that the fibrin gradually forms and webs the blood cells to form a clot, and the increase in platelet-fibrin co-localization indicates the enhanced binding strength of platelets and fibrin. And it can clearly be seen that the bubble stretches along with the coagulation process. We then performed regional fluorescent analysis (white dotted line frame in Figure 3A) around the bubble. The results (Figure 3D) show that the 5, 15, 25, and 35 min average platelet fluorescence intensities around the bubble are 77.870 (σ = 6.988), 129.122 (σ = 8.325), 141.056 (σ = 9.435), and 144.190 (σ = 9.876), respectively, and the corresponding average fibrin fluorescence intensities around bubble are 13.122 (σ = 3.876), 71.645 (σ = 4.844), 83.504 (σ = 5.127), and 85.027 (σ = 5.797), respectively. The fluorescence intensity analysis indicates the increasing strength of the clot around the bubble. The 5, 15, 25, and 35 min average platelet fluorescent areas around the bubble (Figure 3E) are 2.939 (σ = 0.170), 5.689 (σ = 0.278), 7.716 (σ = 0.320), and 7.775 (σ = 0.358) 10³ μm², respectively. The average fibrin fluorescent areas around the bubble are 0.097 (σ = 0.029), 4.367 (σ = 0.130), 5.338 (σ = 0.210), and 5.568 (σ = 0.230) 10³ μm², respectively. The fluorescence area analysis indicates an increasing blood clotting area around the bubble. As expected, the air bubble passively stretches as coagulation progresses, and when the clot tends to be stable, there are insignificant changes of bubble behaviors.

Figure 3.

On-chip adaptive-bubble behaviors during blood coagulation

(A) The fluorescent images of on-chip adaptive-bubble behaviors during blood coagulation. The white dotted line frame refers to the regional area around bubbles. Scale bar: 80 μm.

(B) The three-dimensional fluorescence intensity analysis of platelets and fibrin during on-chip coagulation.

(C) The co-localization analysis of fibrin and platelets during on-chip coagulation.

(D) The mean fluorescence intensity of fibrin and platelets around bubbles (n = 5). Error bars represent standard deviation.

(E) The total fluorescence area of fibrin and platelets around bubbles (n = 5). Error bars represent standard deviation.

Diagnosis model development

Figure 4A shows the user interface and acquisition process of bubble array behavior properties. An integrated imaging system enables continuous image capture of bubble array behaviors during coagulation (10 s per frame), and image cropping is performed to eliminate the interference caused by light scattering from the walls of microfluidic channels in the imaging analysis.^18,20 The developed image algorithms^16,18,20 are used to obtain the adaptive behavior properties of the bubble array (Figure 4B). The acquired adaptive-bubble behavior properties are then used for curve plotting and clinical viscoelastic hemostatic indicator (R, K, Angle, MA, LY30) diagnoses. The included indicators are key parameters in hemostatic diagnosis and cover assessment of coagulation factor levels, fibrinogen function, platelet function, hyperfibrinolysis, etc., and these indicators can be used in combination to map to different medical conditions.^1,2,3 An example curve of the adaptive-bubble array behavior properties of a patient has been shown in the Figure 4C.

Figure 4.

Automated bubble array behavior properties analysis

(A) Imaging processing for bubble array behavior properties acquisition.

(B) The continuous bubble array behavior properties’ acquisition during blood coagulation process.

(C) A curve of adaptive-bubble array behavior properties of a patient.

In this study, 273 clinical patients were collected to develop diagnosis models for clinical indicators (R, K, Angle, MA, LY30) and for the smart diagnoses of hypercoagulation, healthy, and hypocoagulation. The clinical R is used to assess the activity of coagulation factors; it refers to the time required from the clotting activation to the start of fibrin formation, and the healthy R value (TEG analyzer) ranges from 4 to 9 min in clinical diagnoses.^4,5 As shown in Figure 5A, the Passing-Bablok regression analysis of the R value analysis between the two methods shows an A intercept value of −0.4913 with a confidence interval of −0.6727 to −0.2588 and a B slope value of 0.8696 with a confidence interval of 0.8235–0.9091. The Bland-Altman analysis comparing the R value tests between the TEG analyzer and this system shows a mean bias of 1.4267 min with an SD of 0.4674 min (Figure 5C). The limits of agreement (LOAs) range from 0.5107 to 2.343 min. These statistical analyses show that there is a good correspondence for R value tests between this method and the clinical TEG analyzer. We then performed time interval segmentation to establish the association of R value ranges tested via this system with clinical R-indicator diagnosis, as shown in Figure 5E, and the diagnosis model realized 100% diagnosis accuracy (n = 273) of clinical R-indicators as healthy interval ranges from 3 to 7 min.

Figure 5.

Diagnosis interval determination for viscoelastic hemostatic assay indicators (R and K)

(A and B) Passing-Bablok analysis to compare the R- and K-value tests as measured by the clinical TEG instrument and this system (n = 273). The solid red line represents the regression line, and two dotted red lines represent the confidence band in the Passing-Bablok figures.

(C and D) Bland-Altman analysis to compare the R- and K-value tests obtained by the clinical TEG instrument and this system (n = 273). The red solid line is the mean difference of the methods, and the two red dotted lines represent the 95% LOAs.

(E and F) Diagnosis interval segmentation to establish a link of R- and K-value diagnoses between this system and clinical TEG instrument (n = 273).

The clinical K indicator is used to assess the levels and function of fibrinogen; it refers to the time required from the R-time endpoint to a tracing amplitude of 20 mm, and the healthy K-value (TEG analyzer) ranges from 1 to 3 min in clinical diagnosis.^4,5 In this system, the bubble stretching area is used to correspond to the probe amplitude changes of the TEG analyzer, and the bubble stretching area interval, ranging from 0 to 1.0 (102 μm²), is selected to correspond to the tracing amplitude changes from 0 to 20 mm due to the best diagnostic performance (Figure S3). The Passing-Bablok regression analysis (Figure 5B) was performed, and the results show an A intercept value of −0.5000 with a confidence interval of −0.5800 to −0.3938 and a B slope value of 1.5000 with a confidence interval of 1.4375–1.5500. The Bland-Altman analysis comparing the K-value tests between the TEG analyzer and this system shows a mean bias of −0.4718 min with an SD of 0.3864 min (Figure 5D). The LOAs range from −1.229 to 0.2856. These statistical analyses show that this method corresponds well to the clinical TEG analyzer of the K-value test. Time interval segmentation was then performed, and the healthy K-value interval of this method ranges from 1.1 to 3.9 min (Figure 5F; 99.6% diagnosis accuracy, n = 273);

The clinical Angle indicator is used to assess fibrinogen level and function; it refers to the angle value between the horizontal line and the tangent line (from the clot formation point to the maximum arc of the curve), and the healthy Angle value (TEG analyzer) ranges from 53° to 72° in clinical diagnosis.^4,5 As shown in Figures 6A and 6C, Passing-Bablok regression analysis and Bland-Altman analysis were performed, and the results show good agreement between the TEG analyzer and this method for the Angle-value test. The healthy Angle-value interval of this method ranges from 53° to 71° (Figure 6E; 98.9% diagnosis accuracy, n = 273). The clinical MA indicator reflects the strength of the blood clot, which is used to assess platelet function. It refers to the maximum distance between the curves, and the healthy MA value (TEG analyzer) ranges from 50 to 70 mm in clinical diagnosis.^4,5 In this system, the maximum bubble stretching area (MBSA) is used to correspond to the maximum amplitude differences of the TEG analyzer. The Passing-Bablok regression analysis (Figure 6B) was performed, and the results show an A intercept value of −8.1171 with a confidence interval of −8.8234 to −7.5522 and a B slope value of 0.2696 with a confidence interval of 0.2604–0.2810. The Bland-Altman analysis comparing the two methods shows a mean bias of 52.70 with an SD of 4.746 (Figure 6D). The LOAs range from 43.40 to 62.00. These statistical analyses show a good corresponding connection between two methods. We then performed MBSA interval segmentation; the healthy MBSA-value interval of this method ranges from 5.6 to 10.7 × 10² μm2 (Figure 6F; 99.6% diagnosis accuracy, n = 273);

Figure 6.

Diagnosis interval determination for viscoelastic hemostatic assay indicators (Angle and MA)

(A and B) Passing-Bablok analysis to compare the angle value and maximum clot strength as measured by the clinical TEG instrument and this system (n = 273). The solid red line represents the regression line, and two dotted red lines represent the confidence band in the Passing-Bablok figures.

(C and D) Bland-Altman analysis to compare the angle value and maximum clot strength obtained by the clinical TEG instrument and this system (n = 273). The red solid line is the mean difference of the methods, and the two red dotted lines represent the 95% LOAs.

(E and F) Diagnosis interval segmentation to establish a link of the angle value and maximum clot strength diagnosis between this system and clinical TEG instrument (n = 273).

The clinical LY30 indicator is used to diagnose hyperfibrinolysis; it refers to the percentage of the blood clot dissolved at 30 min after the MA value is determined, and the healthy LY30 value (TEG analyzer) ranges from 0% to 8% in clinical diagnosis.^4,5 The Passing-Bablok regression analysis (Figure 7A) was performed, and the results show an A intercept value of 0.0000 and a B slope value of 1.3333 with a confidence interval of 1.2821–1.5000. The Bland-Altman analysis comparing the LY30-value test between the TEG analyzer and this system shows a mean bias of −0.043 with an SD of 0.362 (Figure S10). The LOAs range from −0.753 to 0.667. These statistical analyses show that there is a good agreement for the LY30 test between the two methods. Interval segmentation was then performed, and the healthy LY30-value interval of this method ranges from 0% to 10.1% (Figure 7B; 100% diagnosis accuracy, n = 273). Table S1 summarizes the detailed data from all 273 patients.

Figure 7.

Diagnosis interval determination for viscoelastic hemostatic assay indicator (LY30) and clinical assays

(A) Passing-Bablok analysis to compare the LY30 value as measured by the clinical TEG instrument and this system (n = 273). The solid red line represents the regression line, and two dotted red lines represent the confidence band in the Passing-Bablok figures.

(B) Interval segmentation to establish a link of the LY30-value diagnosis between this system and clinical TEG instrument (n = 273).

(C and D) Heatmap of the viscoelastic hemostatic assay indicators (R value, K value, Angle, MBSA) among patients with hypercoagulation, healthy patients, and patients with hypocoagulation (C), and the three-dimensional tSNE plots (D).

(E) The bubble map of detailed viscoelastic hemostatic assay indicators between the clinical TEG instrument and this system. Red boxes refer to inconsistent diagnostic indicators.

(F) Confusion matrices comparing the performance on 44 clinical patients’ comprehensive coagulation capability diagnoses between TEG instrument and this system.

The clinical coagulation composite index (CI) describes the comprehensive coagulation status of the patient, which includes patients with hypercoagulation, healthy patients, and patients with hypocoagulation. Current clinical instruments are mainly based on empirical algorithms for assessment and quantification.³⁶ Here, artificial intelligence is introduced to achieve a “one-step” smart diagnosis. As shown in Figure S4, the acquired indicators (R value, K value, Angle, MBSA) from adaptive-bubble behavior properties were imported into the developed neural network for comprehensive coagulation capability diagnosis.^16,18,20 The 273 clinical patients’ blood samples (166 healthy patients, CI value: −3 to 3; 63 patients with hypocoagulation, CI value: < −3; and 44 patients with hypercoagulation, CI value: >3) were set for neural network diagnostic model development. As shown in Figure S5, it is clearly distinct in phenotypic space visualized by t-distributed stochastic neighbor embedding (tSNE) between healthy patients, patients with hypocoagulation, and patients with hypercoagulation, and the results show that the developed model possesses a diagnostic accuracy of 97.3% (n = 273).

Clinical assays

Following the preclinical validation studies and model development, we proceeded to patient-oriented testing. We designed a prospective trial with 44 patients in Zhongnan Hospital to determine the usability and accuracy of this system against a standard clinical TEG analyzer. All patients gave informed consent for extra tests. The cluster analysis heatmap (Figure 7C) presents the differences in R value, K value, Angle, and MBSA based on this system between patients with hypercoagulation, healthy patients, and patients with hypocoagulation, and the three-dimensional tSNE plots are shown in Figure 7D. As expected, the method has good discrimination for different coagulation capabilities, and it is clearly distinct in phenotypic space visualized by tSNE. Figure 7E shows the detailed key clinical indicator (R, K, Angle, MA, LY30) diagnoses between this system and the TEG analyzer; the result shows that it realizes 99.1% diagnosis accuracy of the key clinical indicators (n = 220) including R, K, Angle, MA, and LY30 and 100% accuracy (Figure 7F) for hypercoagulation, healthy, and hypocoagulation diagnoses (n = 44). Table S2 summarizes the detailed data from the 44 patients.

Discussion

Herein, we explore and validate a pattern using self-induced adaptive-bubble behavioral perception to freshen the paradigm of viscoelastic hemostatic assays. The adaptive deformation of soft bubbles precisely described viscoelastic hemostatic properties’ transformation during coagulation without the need for active/complex source devices. The bubble array chip allows microassays (≈10 μL blood) and enables multi-bubble tests with good reproducibility, and the developed smart bubble behavior identification empowers automated and user-friendly viscoelastic hemostatic diagnoses. In a prospective clinical study (model development [n = 273]; clinical assay [n = 44]), it exhibits 99.1% diagnosis accuracy of key viscoelastic hemostatic assay indicators (R, K, Angle, MA, LY30; n = 220) and 100% accuracy for smart diagnosis of hypercoagulation, healthy, and hypocoagulation (n = 44). The results support clinical equivalency in accuracy.

Clinical approaches to allow viscoelastic hemostatic assays have evolved from the original TEG 5000 (Haemonetics) and ROTEM (TEM International) to the more advanced TEG 6s and ROTEM Sigma over decades of development. Thanks to the efforts of researchers globally, the technologies are undergoing upgrades, including in resonance frequency, sonic estimation of elasticity via resonance (SEER), ultra-sonic deformation, and laser speckle rheometry.^37,38,39 The clinical performance also validates the feasibility of these approaches. But these methods often require precise active physical source modules or sophisticated optical systems, which also contribute to their limited portability and caseload abilities.^{22,40,41,42,43,44,45,46,47,48,49,50} This work develops a methodology using self-induced adaptive-bubble behavioral recognition to enable hemostatic diagnostics, employing adaptive-bubble behaviors to describe dynamical coagulation processes. The specialized microfluidic bubble platform was designed for adaptive-bubble behavior acquisition, and the developed image/data analysis methods aimed to perform bubble behavior analysis and assist with diagnosis. Prospective clinical studies validate the feasibility and its benefits: it allows easy operation, microassays, and automated diagnosis with excellent portability. This method should enable large-scale and time-sensitive assays in diverse clinical settings including low-resource settings and emergency situations and should contribute to adaptive behavioral perception applications in healthcare, expanding its applicability to clinical and interdisciplinary medicine.

Limitations of the study

The main limitation of our study is the current inability to further explore the scalability of this approach. According to the Young-Laplace equation, there is a potential possibility to realize the viscoelastic hemostasis assays of patients with ultra-hypocoagulation and ultra-hypercoagulation through on-demand size regulation of bubble arrays. Since there are no such special cases among the participating clinical patients, we could not develop and validate the corresponding diagnostic model, but we believe that clinical viscoelastic hemostasis assays for such ultra-hypocoagulable/-hypercoagulable patients will be brought into reality in the near future with the continuous increase and accumulation of clinical cases.

STAR★Methods

Key resources table

REAGENT or RESOURCE	SOURCE	IDENTIFIER
Chemicals, peptides, and recombinant proteins
Calcium chloride solution	Sigma-Aldrich	Cas# 10043-52-4
PE anti-human CD41 Antibody	BioLegend	Cat# 303706; RRID:AB_314376
Alexa Fluor™ 488 fibrinogen	Thermo Fisher	Cat# F13191
Critical commercial assays
Standard clinical Thromboelastography test	Zhongnan Hospital of Wuhan University, Wuhan, China	N/A
Deposited data
The 273 volunteered clinical patients' pathology report (for establishing connections with clinical indicators and the development of intelligent diagnosis model)	This paper	Figshare Data: https://doi.org/10.6084/m9.figshare.24151704
The 44 volunteered clinical patients' pathology report (for double-blind comparison)	This paper
Deep learning model for comprehensive coagulation capacity diagnosis	This paper	Figshare Data: https://doi.org/10.6084/m9.figshare.24151644
Software and algorithms
Python	Python Software Foundation	V3.6.1
GraphPad Prism	GraphPad	https://www.graphpad.com/
ImageJ	NIH	https://ImageJ.nih.gov/ij/
Solid works	Dassault Systemes	https://www.solidworks.com/zh-hans
MATLAB	MathWorks	https://ww2.mathworks.cn/products/MATLAB.html
Android Studio	Google	https://developer.android.google.cn/studio/
3D Studio Max	Autodesk	https://www.autodesk.com.cn/
Other
Lithography and microbubble chip preparation	Key Laboratory of Artificial Micro- and Nano- Structures of Ministry of Education, School of Physics & technology, Wuhan University	N/A
Equipment design and 3D printing	Key Laboratory of Artificial Micro- and Nano- Structures of Ministry of Education, School of Physics & technology, Wuhan University	N/A

Resource availability

Lead contact

Requests for further information and reagents may be addressed to the corresponding author: Yi Yang (yangyiys@whu.edu.cn).

Materials availability

This study did not generate new unique reagents.

Experimental model and subject details

In this work, the 273 clinical patients were collected to develop diagnosis models for clinical indicators (R, K, Angle, MA, LY30) diagnosis, and the smart diagnosis of hypercoagulation, health and hypocoagulation; The method was clinical validated in human subjects (n = 44) for usability and accuracy against a standard clinical device. The all blood samples of clinical patients were collected from the Zhongnan Hospital of Wuhan University. And the volunteers had given informed, written consent. This project (no. 202265K) has been approved and oversighted by Ethical Approval for Clinical/Scientific Research under Medical Ethics Committee, Zhongnan Hospital of Wuhan University. All experiments were conducted in accordance with the Declaration of Helsinki and the International Ethical Guidelines for Biomedical Research Involving Human Subjects.

Method details

System design

The dimension of this device (Figure S1) is 160 mm (L) × 110 mm (W) × 122 mm (H) and the overall weight is 1.1 kg. The device is powered by a lithium battery pack; it can run continuously for approximately 10 h. This device is equipped with: (1) a portable imaging system: a camera module (CMOS IMX214, RERVISION, China), optics lens (KENWEIJIESI, China), USB module (TELESKY, China) and LED module (C5050, BDQ, China). The resolution is 1 μm, the field of view (FOV) is 1.81 mm × 1.02 mm and working distance is 0.75 mm; (2) a portable and miniaturized temperature control system: a polyimide heating film (12 V, 3 W, 48 Ω, Wuxi Tianbo Electrical Appliance Manufacturing Co., Ltd., China) is designed to keep the constant temperature (30°C) on chip. The heated film measures 71 mm (L) x 60 mm (W) with a 26 mm (L) x 25 mm (W) hollow area in the middle for imaging. The commercial temperature controller (XY-WT03-W, OUCHEN, China; accuracy: 0.1°C) is applied to control the temperature of the heating film, which includes a circuit board, an NTC 10K temperature probe and a WIFI module; (3) touchable operation screen and Raspberry Pi module: a capacitive touch TFT screen (resolution: 800 × 480) offers a user-friendly interface for measurement. The Raspberry Pi 4 Computer Model B enables data processing and computing, and the WIFI modular enables cloud based intelligent diagnosis and pathology results sharing. The developed neural network can also run on the Raspberry Pi platform in low-resource areas without network connection; (4) a microfluidic chip with bubble array for viscoelastic hemostatic assays, and a coverslip was put on the surface of chip for oil sealing to avoid evaporation; (5) 3D-ptinted shell design: the shell was fabricated by a 3D printing instrument (ZRapid iSLA660, China) with acrylonitrile butadiene styrene (C-UV 9400E). The interior of the device is divided into several working areas by means of baffles, such as integrated imaging system, lithium battery, Raspberry Pi module, temperature control system module, microfluidic chip module.

Fabrication of the microfluidic chip

The microfluidic chips were made of polydimethylsiloxane (PDMS) prepolymer (SYLGARD 184, Dow Corning, USA), which were fabricated using the standard soft lithography technique. The chip structure was etched onto a silicon wafer (5 inches) covered with the negative photoresist (SU-8 2050, Micro-Chem, USA) via the mask plate (Beijing Machinery Industry Automation Research Institute Co., Beijing, China). The microfluidic chip can be manufactured repeatedly by pouring 30 g PDMS over the master plate and stored in an oven at 75°C for 1 h. The PDMS replica stripped from the master plate is sealed to a slide by plasma instrument (PDC-002, HARRICK PLASMA, USA) for the desired microchannel structures. The microfluidic chip (30 mm (L) × 14 mm (W) × 3 mm (H)) includes an inlet, an outlet, a microfluidic flow channel (200μm (W) × 100 μm (H)) and the bubble array (300 μm (L) × 70 μm (W) × 100 μm (H)).

Temperature selection

Temperature has an effect on the process of bubble behaviors. According to the Ideal Gas Law equation,^9,10,51,52 the negative pressure produced by the clotting process would induce the bubble stretching. Before the point that the bubble stretches to R = D/2 (R: curvature radius; D: bubble column width), bubble curvature (K) continuously increases until the pressure difference is balanced with the Young–Laplace pressure, and the bubble would be stable in this stage.^{9,10,51,52,53,54,55} It’s maximum value as: K = 2/D. Beyond this point, the trend of curvature K changes from previous increase to decreasing, resulting in less Young-Laplace pressure, it causes the bubble to undergo continuous growth.^8,9,51 As expected with our preliminary experiments (Figure S7), we tested a temperature range of 26°C–37°C (spanning from room temperature coagulation test to clinical coagulation test temperature) with a 70 μm of bubble column width, excessive temperature would cause bubbles to stretch beyond the critical point and undergo destabilizing growth, and the 30°C was set due to the balanced bubble behaviors and stability.

Edge computing and deep learning

A touchscreen displays (Raspberry PI Foundation, China) offers a user-friendly interface. Through the portable imaging system,^56,57,58 the continuous images of the adaptive bubble array behaviors are collected, and the interval time was 10s. The image data of adaptive bubble array behaviors are then transferred and analyzed in a Raspberry Pi 4.0 (Broadcom BCM2711 SoC, UK) running Linux. The images are cropped and converted into a grayscale image (8-bit), the light intensity threshold setting is then performed, and the pixel analysis^13,16,17,18 is used to quantify and calculate the adaptive bubble array behaviors properties. The adaptive bubble array behavior properties are then used for curve plotting with Local Weighted Scatterplot Smoothing (LOWESS) processing (10 points in smoothing window),^58,59,60 and clinical indicators (R, K, Angle, MA, LY30) diagnoses. Specially, the R value refers to the time required from the clotting activation to the start of fibrin formation (curve begins to grow); K-value refers to the time required from R-time endpoint to a bubble stretching area of 10² μm²; Angle-value refers to the angle value between the horizontal line and the tangent line (from clot formation point to the maximum arc of the curve); MA-value refers to the maximum bubble stretching area; LY30 refers to the percentage of altered bubble stretching area at 30 min after the MA value is determined. The 273 clinical patients (166 health, CI value: -3-3; 63 hypocoagulation, CI value: <-3; and 44 hypercoagulation, CI value: >3) were set for the development of smart comprehensive coagulation capacity diagnosis. The clinical indicators (R, K, Angle, MBSA) acquired from adaptive bubble behavior properties’ curve were inputted as vector tables for model development, and the trained neural network contained three fully connected layers, the three fully connected layers contained 4, 36 and 9 vectors. The dropout method was used in the training process to avoid overfitting and increase generalization performance. The training set accounts for 70%, and the test set accounts for 30%.^18,20,24 The 70% of the data is randomly selected as the training set each time and the remaining 30% is retained as the validation set during clinical sample training process. The training process performs iterations of twenty cycles. We used a server with a Nvidia Tesla V100 GPU to run the training process over the entire training set. All code was written using Python 3.6.1.

Fluorescent staining of clot

The reagents used in this study includes fibrinogen fluorescence dye (Alexa Fluor 488 human fibrinogen conjugates (F-13191), Life Technologies, USA) and platelets fluorescence dye (PE anti-human CD41 Antibody, Biolegend, China).^34,61 A 1.5 mg/mL fibrinogen fluorescence dye stock solution was reconstituted by dissolving 5 mg of conjugate in 3.3 mL of 0.1 M sodium bicarbonate (NaHCO₃) (PH 8.3) at room temperature, and complete solubilization was realized by taking an hour or more with occasional gentle mixing. The working solution was prepared by adding 100 μL fibrinogen dye to 6 mL distilled water. The undiluted platelets fluorescence dye was stored between 2°C and 8°C, and the use of this reagent was 5 μL per 100 μL of whole blood. In a dark environment at room temperature, the 100 μL of blood sample was incubated (15 min) with 5 μL platelets fluorescence dye (0.2 mg/mL) and 0.13 μL fibrinogen fluorescence dye (2 μg/mL) for dyeing.^16,33 The fluorescent images during coagulation were captured and recorded via a confocal microscope (Nikon, A1R, Japan).

Standard clinical viscoelastic hemostatic assays

The standard clinical viscoelastic hemostatic assays were performed by clinical TEG instrument, and it was performed jointly using two senior laboratory physicians of Zhongnan hospital.

Test stability of this system by different users

We then perform test stability of this system by different users. Three volunteers participated in this test, among them, User 1, User 2 are students with basic medical training and User 3 is a well-trained senior laboratory physician. They performed the TEG tests of a same patient based the clinical TEG instrument and this system, respectively.^62,63 As shown in Figure S6, there are no significant differences on this system with 100% accuracy of the indicators diagnosis. It’s clear to see that this system enables stable test with different operators.

Quantification and statistical analysis

Passing-Bablok regression analysis and coefficient of variation analysis were performed using MedCalc 19.0.7. Bland-Altman analysis was performed using GraphPad Prism version 8.

Published: October 24, 2023

Data and code availability

The confidential medical records data reported in this study cannot be deposited in a public repository. To request access, contact the lead contact. In addition, pathological distribution from these patients and origin code of developed model have been deposited in figshare database: https://doi.org/10.6084/m9.figshare.24151704; https://doi.org/10.6084/m9.figshare.24151644. Any additional information required to reanalyze the data reported in this work paper is available from the lead contact upon request.