An automated, digital immunoassay on a microfluidic cartridge for on-demand cytokine profiling

Abstract

Patients with critical illness often exhibit profound biological heterogeneity, complicating the identification of effective interventions. Resolving distinct molecular profiles to enable timely treatment decisions remains challenging, as integrating biomarker-guided care into routine monitoring is often hindered by fragmented, batch-based workflows. These manual operations decouple molecular data from the acute clinical timeline and are a barrier to reliable, real-time, near to patient, multi-center implementation. To address this gap, we developed an integrated microfluidic digital immunoassay system that achieves high analytical fidelity through a fully automated, simple workflow. The system utilizes a monolithic disposable cartridge to automate bead-based analyte capture, oil-phase partitioning, and signal amplification, eliminating the manual handling and emulsion steps that typically compromise digital assay robustness. The platform enables protein measurement within 45 minutes, achieving sub-picogram limit of detection ( < 0.13 pg/mL), a dynamic range spanning three orders of magnitude, and strong analytical reproducibility. We demonstrate the clinical utility of the system by profiling a validated panel of inflammatory biomarkers in plasma from critically ill pediatric patients, using low sample volumes. Results show strong agreement with gold-standard multiplex assays (R² = 0.925-0.979). By providing a scalable framework for high-fidelity molecular profiling, this system supports the broader goal of accessible, multi-center biomarker validation and the practical implementation of precision medicine in critical care.

Introduction

Precision medicine seeks to advance beyond broad syndrome-based classifications toward individualized treatment strategies grounded in quantitative molecular pathophysiology^1–3. However, implementing this personalized approach in critical care medicine is challenging. Critical illnesses such as sepsis, traumatic brain injury (TBI), acute kidney injury (AKI), acute respiratory failure (ARF), and its most severe form acute respiratory distress syndrome (ARDS) are characterized by profound biological heterogeneity and narrow therapeutic windows, where the opportunity for effective intervention is often measured in minutes to hours^3–5. Importantly, patients presenting with identical clinical syndromes (aka phenotypes) can often be grouped into subphenotypes, which exhibit distinct underlying immune profiles and divergent responses to therapy^6–10. Although circulating protein biomarkers have been identified to stratify these patients^9,11–13, their clinical validation and implementation remain limited by the need for timely, quantitative molecular measurements near patients in the intensive care unit (ICU). Translating biomarker discoveries in critical care into routine clinical research in a multi-center setting for validation and future testing of therapies requires technologies that operate fundamentally differently from those used in discovery-phase proteomics^14,15. In the translational setting, the priority shifts from broad screening to the precise quantification of a focused set of biomarkers in temporal alignment with the patient’s clinical state, enabling real-time subphenotype assignment and evidence-based intervention^15–17. Rather than throughput or panel breadth, the defining requirements in this context are rapid turnaround, analytical fidelity, and the automated robustness required for consistent performance across clinical sites with assays that can be performed on-demand, and near to patient^4,15,17,18.

Existing clinical research workflows are poorly aligned with these requirements, as balancing analytical fidelity with clinical immediacy remains a significant technical challenge. Biomarker measurements are often performed retrospectively in centralized laboratories using batch-based protocols, decoupling molecular data from the acute clinical timeline. Conversely, current point-of-care assays provide the necessary speed but often lack the analytical fidelity required to resolve the wide variability in concentrations observed in critically ill patients with acute immune dysregulation¹⁸. Furthermore, these methodologies, whether centralized or decentralized, frequently rely on manual handling steps that introduce operator-dependent variability. This lack of automation undermines the cross-site consistency essential for multi-center validation, creating a persistent bottleneck in the delivery of high-fidelity measurements with the robustness and usability required for precision medicine–driven clinical research in critical care.

This translational gap stems from the difficulty of reconciling stringent analytical and operational requirements within a single platform. Conventional enzyme-linked immunosorbent assay (ELISA) and other research-use-only immunoassays remain widely used due to their robustness and familiarity, yet their limited sensitivity, constrained dynamic range, and reliance on centralized laboratories preclude the sub-hour turnaround required for acute intervention. Digital ELISA offers a compelling analytical foundation, providing single-molecule sensitivity, improved linearity, and robustness well suited to quantifying immune mediators that fluctuate by orders of magnitude in critical illness^19–24. While various academic microfluidic platforms have explored adapting digital immunoassays for point-of-care operation, leveraging digital microfluidics or droplet-based systems^25–31, full automation remains uniquely challenging. In these systems, reliability and accessibility are often secondary to peak analytical performance, resulting in complex workflows that still require manual intervention. The core difficulty lies in consolidating precise bead handling, stringent washing to preserve signal-to-noise ratios, and reliable multi-phase fluid control for oil sealing into a single, automated sequence. As a result, digital immunoassays remain largely confined to centralized or highly controlled laboratory environments^32,33, constraining their utility for time-sensitive, multi-center clinical research.

Here, we introduce ADAPT (Automated Digital Assay for Precision Treatment), an integrated microfluidic cartridge system designed to bridge this translational gap by providing automated, near to patient digital immunoassays. By consolidating complex multi-phase fluid handling into a single monolithic cartridge, we eliminate the external hardware dependencies and manual emulsion steps that typically render digital assays too fragile outside of a central laboratory. We demonstrate that this automated approach achieves the analytical fidelity required for immune profiling without the infrastructure of a central laboratory. We further validate the platform’s clinical utility by profiling a clinically established inflammatory biomarker panel⁹ in plasma samples from pediatric patients with ARF and benchmarking performance against gold-standard multiplex assays. Together, these results show that near-patient molecular profiling can reliably measure key plasma biomarkers previously used to identify inflammatory subphenotypes³⁴ within a clinically relevant timeframe, positioning ADAPT as a scalable infrastructure for multi-center biomarker validation and precision medicine research in critical care.

Results

Integrated platform design and workflow

ADAPT is a platform engineered to bridge the gap between central-laboratory sensitivity and rapid, on-demand utility. Unlike our group’s earlier digital immunoassay iterations that required high operational complexity and offered limited scalability^35–37, the current system leverages a modular architecture to fully automate the digital immunoassay workflow. This design shift aims to consolidate the multi-step digital assay into a single sequence compatible with the constraints of a clinical setting. The platform comprises three primary modules (Fig. 1a): a central control unit housing the single-board computer and actuators; a reusable manifold that isolates hardware from biological fluids; and a disposable assembly consisting of the fluidic cartridge and integrated detection chip (Fig. 2a). Functionally, the reusable manifold acts as a bridge, interfacing higher-value instrumentation components, such as syringe pumps and linear actuators, with the low-cost disposable unit.

Fig. 1. Overview of the ADAPT platform and paradigm shift in clinical workflow.

a System architecture of the ADAPT platform, consisting of a disposable microfluidic cartridge with a standard microtiter-plate footprint, a reusable pneumatic manifold, and a central control unit with an integrated user interface. b Comparison of conventional centralized laboratory workflows with the ADAPT-enabled near-patient workflow. In the traditional paradigm, samples are transported to centralized laboratories for batch processing, often resulting in a 1–2-day clinical turnaround time before biomarker readout is available for clinical interpretation. In contrast, ADAPT enables a fully automated, simple workflow at or near the patient, delivering quantitative biomarker results in 45 min. This near-patient operation allows immediate evaluation by clinicians and supports timely phenotype assessment and decision-making within clinically actionable timeframes

Fig. 2. Microfluidic cartridge architecture and automated on-cartridge assay progression.

a Illustration of the cartridge design highlighting the pneumatic valving interface (orange) and the dual-planar fluidic inlets (red). Insets display the pinch-valve mechanism in open/closed states and the vertical stacking of sample (upper) and reagent (lower) inlet planes. b Automated on-cartridge assay workflow. The upper timeline illustrates the sequence of fluidic actuation steps controlled by the system, while the lower panels depict the corresponding assay operations within the microwell arrays, including detection chip priming, analyte capture, reagent delivery, washing, substrate incubation, and final oil sealing

The workflow was engineered to support a linear clinical decision path, moving directly from patient admission to phenotype-driven treatment selection (Fig. 1b). To facilitate rapid deployment, the detection chip is pre-patterned with capture antibody-coated beads and stored until the assay is ready to be initiated. Following a standard blood draw and centrifugation step to isolate plasma, the user loads the cartridge with reagents and patient plasma, secures it onto the manifold and initiates the protocol via a graphical user interface (GUI). The instrument autonomously executes the full immunoassay sequence (Fig. 2b), comprising incubation, washing, secondary antibody labeling, and enzymatic development, followed by standalone automated imaging. This approach minimizes the inter-operator variability inherent in manual workflows and reduces the total assay time to approximately one hour, while further optimizations enabled a demonstrated sample-to-answer turnaround of just 45 minutes. Notably, automation reduced the operator workload from >20 manual pipetting and washing steps to only preloading the cartridge with samples and reagents, which led to a streamlined sample-to-answer workflow.

Hybrid manufacturing for scalability

To enable clinical adoption, the system architecture utilizes a hybrid manufacturing strategy that decouples the high-precision microfabrication requirements of the detection chip from the larger features of the cartridge (Fig. S1). The cartridge dimensions are compatible with physical patterning techniques, allowing for the fast manufacturing of millimeter scale features without the need for lithographic precision across the entire device. The microfluidic detection chip contains the high-density microwell array necessary for digital counting. While the analytical and clinical data in this study were generated using PDMS-based soft-lithography detection chips, we performed a feasibility demonstration transitioning the sensor fabrication to an injection-molded Cyclic Olefin Copolymer (COC) format. We confirmed that preliminary fluidic priming and bead loading dynamics remain comparable between the lithographic prototypes and these mass-production compatible equivalents (Fig. S2), though full analytical characterization of the COC-based system remains a focus for next-generation development.

Conversely, the microfluidic cartridge, which constitutes the bulk of the consumable volume, is manufactured using multilayer laminate technology. This approach utilizes discrete 2D layers of laser-cut PMMA and pressure-sensitive adhesive (PSA) to define fluidic paths, replacing the need for complex cleanroom microfabrication or expensive initial tooling. This tool-less workflow allows for rapid design iteration while remaining fully compatible with scalable manufacturing. The high-precision detection chip is integrated into this scalable cartridge via a medical-grade pressure-sensitive adhesive interface, creating a compliant, leak-proof seal. This hybrid integration balances the single-molecule sensitivity of microfabricated sensors with the economic constraints of disposable diagnostics.

Microfluidic cartridge architecture

The cartridge microfluidics were designed to replicate the sensitivity of our previous work^35–37 while enabling robust automation. The detection chip features eight parallel channels for simultaneous multi-sample measurement. Fluid actuation is driven by a hybrid pressure system: on-cartridge reservoirs are pneumatically controlled to drive sample loading, while off-cartridge bulk reagents are delivered via peristaltic pumps. To ensure robust pneumatic coupling without leakage, the cartridge-to-manifold interface utilizes compression-sealed O-rings that engage automatically when the user secures the cartridge.

Flow control is managed by a custom valving interface comprising six on-cartridge pinch valves. As shown in the extracted views in Fig. 2a linear actuator drives a pinch adapter to deform a compliant elastomeric layer, obstructing or enabling flow. To enable the sequential loading of unique samples followed by shared reagents, we implemented a “dual-planar” fluidic architecture; samples enter from an upper inlet plane, while shared reagents enter from a lower plane utilizing a flow-splitting geometry. Figure 2b maps fluidic actuation (top plane) directly to the corresponding molecular events within the microwells (bottom plane), spanning bead priming to immunocomplex formation. This design was favored over individual pinch-valve-controlled inlets to maintain a compact, microtiter-plate-sized footprint. While independent valving could eliminate cross-contamination risks, the added hardware would have introduced significant assembly complexity potentially leading to loading non-uniformities.

To maintain assay integrity, the platform utilizes a multi-stage air-management strategy to address both optical and fluidic risks. A bypass protocol initially purges reagent dead volumes to a waste outlet, supplemented by an integrated hydrophobic membrane vent to release residual air. These measures are critical as air-liquid interfaces can dislodge seated beads or inhibit immunocomplex wetting, which compromises optical clarity. Furthermore, trapped bubbles can create localized resistance imbalances that disrupt loading uniformity and therefore fluidic stability. As a final contingency, the sealing oil phase assists in displacing any air remaining in the cartridge, which is then removed with subsequent wash cycles.

Fluidic characterization and efficiency

Robust automation relies on precise hydraulic control, modeled here via an electrical equivalent circuit to map the system from pressure sources to the on-chip resistance network (Fig. 3a). By analyzing the resistance profiles of the dual-planar inlets, we optimized the hydraulic resistance to ensure active inlet resistance significantly exceeds the outlet path (Fig. 3a, bottom panels). This differential isolates the inactive path during pressurization, preventing cross-contamination. The geometric constraints required to satisfy these resistance ratios were determined using hydrodynamic design-space maps (Fig. S3). Flow uniformity was optimized by characterizing displacement volume and pressure profiles (Fig. 3b, c). While high initial pressurization correlated with non-uniform flow, stability improved drastically after a critical displacement volume was reached. Specifically, dispensing 100 μL for sample reservoirs and 200 μL for the reagent reservoir achieved loading efficiency approaching 100% across all channels. This operating regime ensured uniform delivery without introducing excessive backpressure.

Fig. 3. Fluidic network characterization and performance optimization.

a Hierarchical electrical equivalent circuit model of the fluidic network. Bottom panels display the hydraulic resistance profiles of the dual-planar inlets, demonstrating the resistance differential used to isolate the active inlet from the inactive path. b, c Optimization of displacement volume and actuation pressure for sample and reagent loading, demonstrating uniform flow distribution across all channels under optimized operating conditions. d Characterization of washing efficiency as a function of flow rate, establishing the minimum fluidic exchange volume required for complete reagent removal. e Validation of the on-cartridge mixing unit, showing rapid mixing dynamics achieved through pulsed actuation and confirmed by colorimetric analysis at the mixer outlet

To maximize analytical sensitivity, we quantified wash efficiency (Fig. 3d), determining that 300 μL is sufficient for complete reagent removal and background suppression. While higher flow rates expedited washout, complete exchange was consistently achieved between 100 and 300 μL. This establishes a practical lower bound for wash steps balancing background suppression with reagent conservation. Finally, on-cartridge mixing was optimized to prevent premature catalytic turnover of the chemiluminescent substrate. The unit incorporates a hydrophobic vent to allow air escape during actuation, ensuring complete filling. Three short, high-pressure pulses yielded approximately 95% mixing efficiency within 15 s, as confirmed by optical analysis of dyed solutions upstream of detection chip entry (Fig. 3e). Computational fluid dynamics (CFD) revealed that this efficiency is attributable to a transition from a diffusion-limited laminar expansion to an inertia-driven recirculating regime. At higher inlet velocities, the onset of flow separation generates symmetric transverse vortices that increase the interfacial area available for homogenization through passive stretching and folding. This transition was validated through scaling experiments (Fig. S4), confirming that the recirculating topology achieves higher mixing quality in significantly less time than the stable expansion characteristic of lower flow rates.

Analytical validation

To evaluate the analytical performance of the ADAPT platform, we conducted experiments to characterize sensitivity and specificity for multiplex detection. Figure 4a–c presents titration standard curves for three biomarkers, soluble tumor necrosis factor receptor 1 (sTNFr1), interleukin-8 (IL-8), and interleukin-6 (IL-6) that can be used to distinguish subphenotypes in children with acute respiratory failure with or without PARDS^9,34. Recombinant proteins were spiked into assay buffer (see Methods) at varying concentrations and measured using the fully automated on-cartridge protocol. Each data point represents the average signal from independent replicates, with data fitted using four-parameter logistic (4-PL) regression used to generate the standard curve and to calculate the Limit of Detection (LOD). Quantitative analysis demonstrated high sensitivity, yielding LODs of 0.188, 0.810, and 0.128 pg mL-1, and Limits of Quantification (LOQ) of 25, 22, and 5 pg mL-1 for sTNFr1, IL-8, and IL-6, respectively (Table 1).

Fig. 4. Analytical validation of the ADAPT platform.

a Soluble Tumor Necrosis Factor receptor 1 (sTNFr1), b Interleukin 8 (IL-8), and c Interleukin 6 (IL-6) measured over concentration ranges of 0.128 pg/mL–10000 pg/mL for sTNFr1 and 0.04 pg/mL–3200 pg/mL for IL-8 and IL-6 in ELISA dilution buffer (1% Bovine Serum Albumin). Digital immunoassay signals, expressed as average enzyme per bead (AEB), were fitted using four-parameter logistic (4PL) regression. The black dotted line represents 3σ above the blank signal, which is used to estimate the limit of detection (LOD) for each cytokine, and the black dashed line represents the blank signal. d Assay specificity evaluated using single analyte spike-in, all spike-in, and blank (negative) samples of recombinant cytokine markers at 5000 pg/mL sTNFr1 and 1600 pg/mL IL-8 and IL-6 in ELISA dilution buffer. Dotted lines are showing the blank signal across single and all spike-in signal levels. e Intra-cartridge precision assessed by the coefficient of variation (CV) of IL-8 AEB measurements from four technical replicates within a single channel, evaluated across eight parallel channels. The black dashed line indicates the overall mean AEB across channels. f Inter-cartridge precision evaluated by CV analysis of six standard concentrations measured across four independently fabricated cartridges, demonstrating robust assay reproducibility and automated operation

Table 1.

Nonlinear regression parameters and analytical performance characteristics for sTNFr1, IL-8, and IL-6

Biomarker	LOD (pg/ml)	LOQ (pg/ml)	Dynamic Range (pg/ml)	Intra-CV	Inter-CV
sTNFr-1	0.188	24.97	~25 to 2441	6.25%	7-15%
IL-8	0.810	22.36	~22 to 285		4-21%
IL-6	0.128	5.01	~9 to 1100		3-15%

Data represent best-fit values derived from a four-parameter logistic (4PL) regression model. The limit of detection (LOD) and limit of quantitation (LOQ) were calculated as the mean signal of the assay blank ± 3 and 10 standard deviations (SD), respectively. The dynamic range defines the effective quantitative interval between the LOQ and the saturation limit. Precision is reported as the percent coefficient of variation (%CV) of the Average Enzymes per Bead (AEB) signal. Intra-cartridge variability was determined from four replicates of IL-8 measured across eight independent channels within a single device. Inter-cartridge variability was assessed across four separate cartridges using six standard concentrations to demonstrate fabrication and automation robustness

To assess assay specificity, we evaluated single-spike-in versus all-spike-in profiles to quantify off-target interference (Fig. 4d). Results demonstrated negligible antibody cross-reactivity across all markers, as signals generated by off-target analytes were comparable to background levels. Furthermore, comparison of single-plex versus 3-plex configurations confirmed robust multiplexing capability; each analyte maintained a clear signal-to-blank ratio in the multiplexed state, with minimal interference or signal compression observed in the presence of non-target proteins.

To evaluate the reproducibility of our mass-production compatible cartridge fabrication, we characterized system precision. The intra-cartridge CV (variation between channels) was 6.25%, specifically confirming loading uniformity and precise positive pressure control across parallel channels. Meanwhile the inter-cartridge CV (variation across four different devices) ranged from 3% to 21% across the biomarker panel (Fig. 4e, f and Table 1). While this inter-cartridge variance is likely caused by manual assembly and lamination, these results demonstrate that the manufacturing process maintains acceptable performance across independent devices.

Clinical utility validation

To validate the platform in a relevant clinical context, we benchmarked ADAPT against the pathophysiological interquartile ranges (IQRs) reported for sTNFR1, IL-8 and IL-6 in studies using biomarkers for subphenotyping adults with ARDS⁶ and children with ARF and/or ARDS⁹. While the system maintains a linear dynamic range up to the Upper Limit of Quantification (ULOQ) (2441, 285, and 1100 pg mL^-1 for sTNFr1, IL-8, and IL-6, respectively), non-linear regression (R² = 0.99) extends the quantifiable range to saturation limits of 10,000 pg mL^-1 (sTNFr1) and 3200 pg mL^-1 (IL-8/IL-6) (Fig. 5a). Unlike standard assays limited to smaller linear regions, this extended range encompasses the full IQR of biomarker concentrations associated with these clinical cohorts, enabling rapid quantification across the physiological biomarker ranges without the need for additional dilutions and assay repeats.

Fig. 5. Clinical validation using pediatric ARDS patient samples.

a Comparison of the linear dynamic range and saturation limits of the ADAPT platform with pathophysiologically relevant cytokine concentration interquartile ranges (IQRs) reported for adult ARDS cohorts (Calfee et al.⁶) and pediatric ARDS subphenotypes (Dahmer et al.⁹). b–d Method comparison between ADAPT and the gold-standard Luminex assays for sTNFr1, IL-8, and IL-6, respectively, using plasma samples from pediatric ARDS patients (n = 21 for sTNFr1 and n = 20 for IL-8 and IL-6). Strong correlation was observed between the two methods (R² = 0.9790 for sTNFr1, 0.9245 for IL-8, and 0.9612 for IL-6). The dashed line indicates the line of identity (y = x)

We measured the biomarker panel in patient samples, with paired measurements available in 20 samples for sTNFr1 and 21 for IL-8 and IL-6. Comparisons to the Luminex platform revealed strong linear correlations for all three markers (R² = 0.98, 0.92, and 0.96 for sTNFr1, IL-8, and IL-6, respectively; Fig. 5b–d). The regression slopes were 0.91 for sTNFr1, 1.39 for IL-8, and 1.43 for IL-6. The slight deviation observed in the lower concentration regime (notably for IL-8 below 20 pg/mL) likely reflects the transitional quantification limits between the two platforms, where sample concentrations approach the ADAPT lower limit of quantitation (LOQ) while simultaneously remaining near the Luminex noise floor.

Discussion

The implementation of precision medicine in the intensive care unit is often constrained by the inability to rapidly assay biomarkers needed for subphenotyping. This classification is integral, enabling predictive enrichment strategies to be implemented in precision-medicine based randomized controlled trials^4,38. ADAPT addresses this challenge by automating single-molecule counting to deliver results within 60 minutes in a benchtop, readily accessible format. By combining sub-picogram sensitivity with a streamlined workflow, the platform also demonstrates that high-fidelity immune profiling is possible without a centralized infrastructure. This capability is particularly relevant for managing heterogeneous syndromes, where the ability to stratify patients within the first few hours of admission may influence the efficacy of potential immunomodulatory interventions.

Recent multi-center studies have demonstrated that many of these heterogeneous syndromes can be divided into distinct hypo- and hyper-inflammatory subphenotypes with divergent clinical outcomes and responses to therapy. Specifically, foundational work by Calfee et al. demonstrated that clinical and biological data, which included IL-6, IL-8, and sTNFr1 among other key biomarkers, could reproducibly differentiate adult patients into two distinct ARDS subphenotypes with divergent clinical outcomes and responses to mechanical ventilation ⁶, intravenous fluid¹² and statin-based treatment⁷. Recently, Dahmer et al. extended these findings to the pediatric population, demonstrating that a parsimonious panel of these same three proteins could effectively differentiate children with inflammatory subphenotypes linked to increased mortality and prolonged mechanical ventilation⁹. Consequently, we focused on these three validated markers. This approach allowed us to optimize the fluidic architecture for speed and quantitative precision rather than comprehensive multiplexing. Our validation results, using plasma from critically ill children with ARF, support this strategy, demonstrating high correlation (R² = 0.925-0.979) with gold-standard methods even in complex plasma matrices.

Beyond biological validation, the system architecture directly addresses the batch processing delays inherent to commercial digital ELISA platforms (e.g., Simoa, Luminex), which typically accumulate samples over 12–24 hours to justify reagent costs. ADAPT’s unitized cartridge design decouples cost from throughput to enable rapid access testing. ADAPT’s flexibility also addresses the limitations of currently available decentralized systems. Commercial systems such as Ella (Bio-Techne)³⁹ and MultiStat (Randox Laboratories)⁴⁰, while automated, are often constrained by rigid, fixed cartridge architectures that may make adapting to novel signatures challenging. In contrast, ADAPT separates the detection surface from fluidic handling, allowing the biomarker panel to be reconfigured by swapping the detection chip insert. This addresses the fixed-panel constraints, enabling more rapid adaptation to emerging clinical evidence. This architecture achieves a technical advance over existing digital platforms by integrating high-sensitivity single-molecule counting within a modular, self-contained fluidic ecosystem designed for single-patient, on-demand sample processing. A detailed quantitative comparison of analytical specifications with existing platforms is provided in Table S1.

Critically, the robust analytical performance required for this profiling was achieved without sacrificing economic viability by adopting a modular engineering strategy that explicitly decouples high-precision functions from the bulk consumable. We mitigated the scalability constraints often encountered in microfluidics by using a hybrid manufacturing approach: micron-scale features are confined to the sensor substrate, while the bulk cartridge is fabricated from commoditized laminates to reduce the cost-per-test. This modular approach enhances fluidic reliability by locating the high-volume wash buffer off-cartridge to enable robust background suppression, a capability often limited in fully integrated devices by volumetric constraints. Furthermore, because the detection module is optically decoupled from the fluidics, the current microscope-based validation can be seamlessly swapped for a miniaturized optical module, demonstrated in previous work³⁷, without requiring consumable redesign, ensuring the platform remains adaptable to component-level optimization.

Despite these advances, the current platform has limitations. From a transport perspective, a positive-pressure, unidirectional flow architecture was selected to maximize stability and reproducibility. While robust, this design inherently places an upper bound on capture efficiency compared to actively recirculating systems. Clinical utility validation in this study was limited to banked, retrospective samples and off-cartridge plasma separation via centrifugation. While this approach enabled rigorous benchmarking against gold-standard assays, from a workflow standpoint, this methodology still precludes point of care operation in an intensive care unit environment. Future work requires validation of the ADAPT system using fresh whole blood and evaluation of real-time subphenotypic stratification to mimic an active ICU workflow.

Subsequent iterations will aim to address these limitations to facilitate routine clinical adoption. To overcome the diffusion-limited regime governed by our current unidirectional flow, we aim to integrate inline valving within the split channels. This addition is intended to provide precise flow control necessary to maintain stable recirculation without introducing instability in positive pressure systems, thereby improving sensitivity via advective mass transport. Simultaneously, to realize a true near to patient capability, we envision coupling the cartridge inlet with a passive vertical filtration unit based on methods described by Li and Steckl⁴¹. In this module, whole blood applied to the filter site separates via capillary action, flowing into a reservoir designed to automatically meter the precise plasma volume required for the assay, a process estimated to add only 5-10 minutes to the total assay time. We have also demonstrated the feasibility of direct whole blood analysis in small animal models in a previous publication³⁷; however, to our knowledge, direct phenotypic/subphenotypic stratification from whole blood has not yet been demonstrated in the clinical literature. Finally, future prospective studies will seek to deploy the system in active ICU workflows, aiming to demonstrate that rapid, on-demand subphenotypic stratification can lead to improved patient outcomes. By connecting academic innovation with scalable manufacturing, ADAPT represents a deployable infrastructure for multi-center biomarker validation with applicability to precision medicine research, bridging the gap between biomarker discovery and actionable clinical insight in critical care medicine.

Materials and methods

Microfluidic detection chip fabrication and patterning

Microfluidic detection chips were fabricated using standard soft lithography and microfabrication techniques. Silicon wafers were patterned with microarray structures using optical lithography followed by deep reactive ion etching (DRIE) to create master molds, which were silanized to facilitate demolding. A thin layer of PDMS (10:1 elastomer to curing agent ratio, Sylgard 184, Dow Corning) was spin-coated onto the mold, thermally cured, then transferred to the glass slide, detailed in our previous publication³⁵. The microarray sensor substrates (Beijing Poly Microchip Technology Co) were fabricated from Cyclic Olefin Copolymer (COC) using high-precision injection molding.

Prior to use, the detection chip was seeded with magnetic microbeads (Dynabeads M-270 Epoxy, Invitrogen) conjugated with capture antibodies. Multiplexed patterning was achieved using a removable PDMS microchannel oriented perpendicular to the assay flow. Diluted bead suspensions were introduced via automated pipetting, allowed to settle, and washed. To prevent air-interface disruption, the patterning mask was removed while the assembly was submerged in a deionized water bath. The detection chip was immediately bonded to the top flow-cell layer using a pressure-sensitive adhesive (PSA), rehydrated, blocked in blocking buffer (SuperBlock™, Thermo Fisher Scientific) for 30 minutes, and stored in storage buffer (PBS pH 7.4 containing 0.1% BSA and 0.05% sodium azide) at 4 °C. This storage protocol was designed to maintain bead stability and binding shelf-life for up to three months, facilitating the use of pre-patterned detection chips for on-demand clinical testing.

Disposable cartridge architecture and assembly

The disposable cartridge was manufactured using a multilayer laminate process (Fig. S1). The device stack consists of laser-cut acrylic layers (McMaster-Carr), alternating with double-sided PSA layers (Adhesives Research). Prior to assembly, the layers were cleaned via sonication in isopropanol and preheated to 60 °C to enhance adhesion. The stack was aligned using dowel pins and compressed in a hydraulic press at 30 MPa to ensure uniform bonding.

Fluid control was achieved via integrated on-cartridge valving. A laser-cut silicone elastomeric membrane (60 A durometer, McMaster-Carr) was bonded to the cartridge underside. To prevent fluid backflow and ensure bubble-free operation, fluidic lines were lined with hydrophobic membranes (Sterlitech), with specific venting membranes placed along shared liquid lines for passive air removal. Assembled cartridges were stored for at least 24 h prior to use.

Instrumentation and cartridge interface

The automated platform is controlled by a Raspberry Pi 3B+ single-board computer running custom Python scripts. The instrument interfaces with the disposable cartridge via a reusable CNC-machined aluminum manifold, which routes pneumatic and fluidic connections through sealed O-ring interfaces. Fluid actuation was driven by a hybrid pressure system: reagents were dispensed from on-cartridge reservoirs using syringe pumps (positive pressure, Tecan XC), while high-volume wash buffer (PBS pH 7.4 containing 0.05% (v/v) Tween-20) was delivered via an off-cartridge peristaltic pump (Welco WPM). Precise flow gating was managed by six miniature linear actuators (Haydon-Kerk Pittman) and solenoid valves (Burkert) for venting control. The experimental setup and instrumentation used for this study are shown in Figure. S5.

Fluidic characterization

To quantify the efficiency of washing, mixing, and reagent loading, we performed fluidic characterization using high-contrast dyes against a bright background behind the detection chip. All sequences were photographically recorded, and the resulting images were processed using ImageJ. Images were converted to 8-bit grayscale, and the mean gray value (intensity) was measured within the regions of interest, such as the mixing chamber or detection channels. To determine efficiency, dye intensity values were normalized on a scale from 0 to 1, where 0 represented the intensity of the pure wash buffer and 1 represented the intensity of the loaded dye. Loading uniformity and washing and mixing efficiencies were calculated by comparing these normalized intensity values across multiple conditions. Uniform loading or effective washing was defined by all eight channels reaching within 5% of the target intensity. Mixing homogeneity was defined by the intensity variance across the mixing chamber, ensuring consistent reagent concentration before delivery to the detection chip.

Assay reagents and automated operation

Reagents for the quantification of human IL-6 (Cat# 501126 and 501201; BioLegend), IL-8, and soluble TNF receptor-1 (sTNFr1) (Cat# DY208 and DY225, DuoSet®; R&D Systems) were prepared using commercial ELISA kits, utilizing biotinylated detection antibodies and a streptavidin–horseradish peroxidase (HRP) conjugate. Assay buffer (Phosphate-buffered saline (PBS) pH 7.4 containing 1% (w/v) Bovine Serum Albumin (BSA) and 0.05% (v/v) Tween-20) was used for sample dilution and reagent preparation. The chemifluorescent signal was generated using the QuantaRed™ enhanced HRP substrate (Thermo Fisher), while fluorinated oil (Novec™ 7500, 3 M) served as the sealing fluid for digital quantification. The user workflow begins by preloading the designated on-cartridge reservoirs with the required reagents and patient plasma samples via micropipette. The cartridge is then sealed by the loading inlets and secured onto the pneumatic manifold using a calibrated torque-limited fastening interface to ensure uniform interfacial pressure. Following this setup, the system is initialized via the custom GUI. To enable precise quantification while accounting for potential inter-run variability for measurements using the ADAPT system, each assay run incorporated an internal four-point standard curve alongside the clinical samples. The concentration of each unknown sample was determined by interpolating its mean signal intensity from this concurrent standard curve.

Building on our previously reported pre-equilibrium digital assay workflow³⁶, the automated assay sequence began with a system priming step using wash buffer to wet the microfluidic network (Figure. S6). Patient plasma (8 µL), diluted 1:10 in assay buffer, was introduced into the detection chip channels under positive pressure. This brief incubation ensured that analyte capture occurred within the initial linear regime of binding kinetics, substantially reducing assay time while preserving high analytical sensitivity³⁶. Unbound analytes were subsequently removed via a high-volume wash step using wash buffer delivered by the off-cartridge peristaltic pump³².

The assay then proceeded with the sequential delivery of detection antibodies followed by the enzyme conjugate. For each step, 800 µL of reagent was distributed across the eight channels. Consistent with the pre-equilibrium nature of the assay, these delivery steps were strictly timed to ensure reproducible, non-equilibrium binding across all sensor arrays. After the final wash to remove unbound enzyme, the two-component chemifluorescent substrate was mixed on-cartridge through a pressurized venting sequence and incubated within the sensor array. Finally, fluorinated oil was introduced to displace the aqueous substrate solution, sealing the femtoliter-volume microwells for endpoint detection. The complete fluidic operation, from sample loading through assay completion, required approximately ~35 minutes.

Imaging and data analysis

Immediately following the assay, the detection chip was imaged using an automated fluorescence microscope equipped with a consumer-grade CMOS camera (Ximea) and a 5x objective. The system scans 48 microarrays in both fluorescence (545/605 nm) and bright-field modes using a previously described automated scanning and auto-focusing algorithm³⁷.

Images were analyzed using an in-house developed Convolutional Neural Network (CNN) based on a U-Net architecture^42,43. Instead of standard intensity integration, the network accepts raw fluorescence micrographs as input to perform semantic segmentation, generating a binary digital output of enzyme-active (“On”) microwells while simultaneously identifying and masking sensor defects. Bright-field images were processed using Sobel edge detection to confirm bead occupancy. The algorithm corrects the data by excluding masked defects and normalizing the “On” count against the valid bead-loaded well count to determine the fraction of active wells (P_on). Finally, P_on was converted to the average enzymes per bead (λ) using Poisson statistics (λ = −ln(1−P_on)).

Clinical sample validation

To evaluate the platform’s performance in patients with critical illness, we accessed residual, de-identified plasma samples from critically ill children enrolled in the University of Michigan Expediting Bedside use of a Novel Microfluidics - Based Platform for the Identification of Subphenotypes in Critically Ill Pediatric Patients (aka Microfluidics) study (University of Michigan IRB HUM00251815). For the Microfluidics study, all included patients were prospectively identified and written, informed consent obtained from a parent or legal guardian. Inclusion criteria included infants > 42 weeks corrected gestational age to children and adolescents <21 years old (the usual age ranges of PICU patients). All patients had a diagnosis of ARF, defined as the need for mechanical ventilation or noninvasive positive pressure ventilation, and/or PARDS, as defined by the Pediatric Acute Lung Injury Consensus Conference (PALICC-2) definition⁴⁴.

Prior to analysis, all samples were de-identified and stored at −80 °C. Concentrations of sTNFR1, IL-6, and IL-8 were determined using custom Human Luminex Discovery Assays (R&D Systems). Specifically, sTNFR1 was measured using a 2-plex panel (Catalog # LXSAHM-02) that included intercellular adhesion molecule 1 (ICAM-1). For this panel, plasma was diluted 1:9 in calibrator diluent RD6-52. IL-8 and IL-6 were measured as part of a 4-plex panel (Catalog # LXSAH-04) that also included angiopoietin-2 (Ang-2) and the receptor for advanced glycation end products (RAGE). For the 4-plex, plasma was diluted 1:3 in the same diluent. All assays were performed according to the manufacturer’s instructions and read on a Bio-Plex MAGPIX Multiplex Reader (Bio-Rad Laboratories).

Statistical analysis

Assay performance was validated using recombinant protein standards spiked into assay buffer. Calibration curves were generated by fitting the mean signal intensities to a four-parameter logistic (4PL) nonlinear regression model using GraphPad Prism. The Limit of Detection (LOD) was calculated as y_blank + 3σ, and the Limit of Quantification (LOQ) as y_blank + 10σ, where σ is the standard deviation of the residuals (S_y.x) from the low-concentration regression. To evaluate analytical specificity, cross-reactivity was assessed by spiking individual analytes and a 3-plex cocktail into the system. Precision was evaluated using the Coefficient of Variation (CV), defined as the ratio of the standard deviation (σ) to the mean (μ) signal intensity (CV = (σ/μ)×100). Intra-assay precision was assessed using replicates within a single cartridge run, while inter-assay precision was determined across independent experiments. Correlation between the ADAPT measurements and the reference Luminex Bio-Plex data (tested at University of Michigan) was determined using linear regression analysis. The goodness of fit (R²) and slope were calculated based on paired clinical data points (n = 21 for sTNFr1 and n = 20 for IL-8 and IL-6) to assess analytical correlation. Clinical validation experiments were performed using two technical replicates per sample due to limited pediatric plasma volume availability. All other analytical characterization studies, including standard curves, inter-assay precision, and cross-reactivity measurements, were conducted using at least four technical replicates per condition (excluding optimization experiments). Intra-assay precision was evaluated using four technical replicates within each channel.

Author contributions

A.D.F. led the study, contributing to conceptualization, investigation, methodology, software, data curation, formal analysis, and writing the original draft. M.N. contributed to methodology and investigation. M.F. contributed to investigation. R.Y. contributed to software. N.H. contributed to investigation and resources. M.K.D. and H.F. contributed to conceptualization, resources, analysis of data, and manuscript review and editing. M.X.T. contributed to conceptualization and methodology. Y.S. contributed to conceptualization, methodology, software, supervision, and writing – review & editing. K.K. contributed to conceptualization, funding acquisition, supervision, and writing – review & editing.

Competing interests

The authors declare no competing interests.

Supplementary information

The online version contains supplementary material available at 10.1038/s41378-026-01374-2.