Confinement-guided ultrasensitive optical assay with artificial intelligence for disease diagnostics

The necessity for ultrasensitive detection is becoming increasingly apparent as it plays a pivotal role in disease early diagnostics and health management, particularly when it comes to detecting and monitoring low-abundance biomarkers or precious samples with tiny volumes. In many disease cases, such as cancer, infectious disease, autoimmune disorder, and neurodegenerative disease, low-abundant target biomarkers like circulating tumor cells (CTCs), extracellular vesicle (EV) subpopulations, and post-translational modified proteins (PTMs) are commonly existing and can be served as early indicators of disease onset or progression. However, these biomarkers often exist in ultra-low quantities in body fluids, surpassing the detection limits of conventional diagnostic tools like enzyme-linked immunosorbent assay (ELISA). This leads to the inability to probe disease evolution at a very early stage from molecular pathology perspective. In such regard, ultrasensitive optical assays have emerged as a solution to overcome these limitations and have witnessed significant progress in recent decades. This review provides a comprehensive overview of the recent advancements in ultrasensitive optical detection for disease diagnostics, particularly focusing on the conjunction of confinement within micro-/nano-structures and signal amplification to generate distinguishable optical readouts. The discussion begins with a meticulous evaluation of the advantages and disadvantages of these ultra-sensitive optical assays. Then, the spotlight is turned towards the implementation of artificial intelligence (AI) algorithms. The ability of AI to process large volumes of visible reporter signal and clinical data has proven invaluable in identifying unique patterns across multi-center cohort samples. Looking forward, the review underscores future advancements in developing convergent biotechnology (BT) and information technology (IT) toolbox, especially optical biosensors for high-throughput biomarker screening, point-of-care (PoC) testing with appropriate algorithms for their clinical translation are highlighted.



INTRODUCTION
Biomarkers discovered and validated from biological fluids hold great potential for reflecting the molecular pathology and progression of diseases. [1][2][3] Compared to imaging techniques, e.g., computed tomography (CT), 4,5 magnetic resonance imaging (MRI) 6,7 which highly rely on well-trained clinician to identify nidus of milli-meter to centimeter dimension, 8,9 biomarkers based in vitro diagnostics (IVD) can offer a range of advantages in disease diagnosis and monitoring, [10][11][12][13] including high sensitivity, 14 real-time monitoring, [15][16][17][18][19] facilitation of personalized treatment, 20,21 and cost-effectiveness. 22,23 However, biomarkers like post-translation modification (PTM) proteins, [24][25][26] microRNA, [27][28][29][30][31] extracellular vesicle (EV) subpopulations, 19,[32][33][34][35][36] circulating tumor cells (CTCs), 37,38 and circulating nucleic acids 39,40 are usually present in low abundance, and their secretion is rare during the early stage of disease progression, posing significant challenges for the conventional detection methodologies like enzyme-linked immunosorbent assays (ELISA) 41,42 due to their limited sensitivity and unachievable stable monitoring. Usually, compromised strategies have to be used, such as consumption and conversion of large input volume into a squeezed detection volume to achieve a decent enrichment factor. This not only increases the economic and psychological burden on patients associated with multiple sampling but also constrains the widespread adoption and implementation of biomarkers in clinical applications. In such regard, the development of ultrasensitive toolboxes is essential with high demand.
The pressing need for ultrasensitive detection methods in the realm of biomarker analysis has led researchers to explore various innovative strategies. Fundamentally, ultrasensitive detection either relies on high-end detection devices or signal amplification strategies. The former category includes advanced instruments such as total internal reflection fluorescence (TIRF) microscopy 43,44 and cutting-edge flow cytometry analyzers. 45,46 While these state-of-the-art tools provide exceptional sensitivity and accuracy, their widespread accessibility may be hindered by factors such as high costs and the need of specialized expertise in operation and maintenance. The latter, on the other hand, serves as the primary focus of this review, offering a more promising and accessible avenue for exploration and translation.
Typically, signal amplification with confinement strategies, e.g., multiple enzymatical catalysis, nucleic acid amplification, and surface plasmonenhanced fluorescence, involve generating sufficiently robust output signals, generally achieved by confining signal molecules within minuscule spaces. Such confinement can be realized either using semi-enclosed microwells, fully enclosed water-in-oil droplets, or by immobilizing ultra-bright or visible probes on the surfaces of particulate substrates to provide localized tremendous reporter concentration and signal readout ( Figure 1). These methodologies can enhance signal detection and improve the overall sensitivity of the method, ultimately pushing the detection limits down to even single-entity level.
In this review, we primarily focus on the development of ultrasensitive optical sensors, emphasizing the "adhesion" of visible reporters to the outer or internal microenvironment, presenting external-and internal-confinement strategies, respectively. Generally, plasmonic [47][48][49][50][51][52] or non-plasmonic materials can be employed to convert target binding events into digitized signal format ( Figure 1). How artificial intelligence facilitates image processing, particle counting, and clinical data handling will be emphasized. We introduce representative cases, discussing their pivotal features and shortcomings. Finally, we delve into the prospective developments both in AI technologies and in the field of ultrasensitive optical detection. Our vision is that the synergy of these advancements will significantly contribute to reshaping diagnostic approaches in healthcare, offering a more precise, efficient, and reliable analysis.
in such system, reporter molecules with high-density would generate localized high signal intensity on the individual beads that can "illuminate" the single-bead microenvironment under excitation, resulting in superior sensitivity. Such "ultrabright" beads can be easily detected using conventional microscope with low-magnification even mobile phone-based microscope. Different feasible algorithms can then be employed for beads digital counting, which can be converted to the target analyte concentration using appropriate algorithm. The screened biomarker can be used for toward separation of different groups in a cohort study to permit disease early diagnostics, stratification, therapy monitoring, etc. [53][54][55][56] So far, plasmonic-and nonplasmonicmaterials have been widely used as optical labels in cooperation with confinement strategy for ultrasensitive assay development, which will be discussed in the following sections.

Nanoplasmonics based reporter
When considering the direct signal output, plasmonic objects with size larger than ~30 nm can be visualized under dark-field microscopy (DFM) due to their strong scattering capabilities. This arises from the localized surface plasmon resonance (LSPR) effect. When the size of the plasmonic particles exceeds a certain threshold-yet remains smaller than the wavelength of light-there is a significant enhancement in light scattering. 48,[57][58][59][60] On the other hand, plasmon can also enhance fluorescence through surface plasmon interactions by mediating the distance (e.g. adjusting the spacer thickness) between the plasmonic materials and the fluorophore. These interactions amplify the local electromagnetic field. This, in turn, increases the excitation rate and the quantum yield of the fluorophores, resulting in ultimately enhanced fluorescence signals. 61,62 These intrinsic advantages permit plasmonics-related detection with achievable ultrasensitivity. The most widely explored formats are plasmonic label by its alone and surface plasmon-enhanced fluorescence.
Nanoplasmonics alone. The surface plasmon effect refers to the collec-tive oscillation of electrons at the interface between a metal and a dielectric material, leading to unique optical phenomena. 49 Dark-field spectroscopy (DFS) is a typical plasmonic particle spectroscopy technique that employs microscopy in combination with a spectrometer to observe plasmonic particles. 63, 64 Lee group 65 developed a plasmonic exosome assay using transmission surface plasmon resonance with antibodies functionalized nanohole arrays to identify ovarian cancer exosomal surface biomarkers ( Figure 2A). They demonstrated improved signal amplification through secondary labeling with nanoprobes such as gold nanostars of variable sizes and dimensions for detection of 12 surface markers. The sensitivity enhancement factors were 10 4 -fold and 10 2 -fold compared to Western blot and chemiluminescence ELISA, respectively. The plasmonic structure sensing mechanism was confirmed through Three-Dimensional Finite-Difference Time-Domain simulations (3D-FDTD) simulations, showing enhancement with the immune-sandwich format on a glass substrate. By using this technique, exosomal EpCAM and CD24 were stably probed, and were demonstrated the ability to differentiate ovarian cancer patients (n=20) from non-cancer patients (n=10).
In another work, Hu and co-workers 66 built a nanoplasmon-enhanced scattering (nPES) assay to directly probe the membrane antigens from tumour-derived EVs, in which the detection antibody-conjugated gold nanospheres (50 nm) and nanorods (25×60 nm) were bind to EVs, which were captured by antibodies pre-decorated glass chip to form the "sandwich" format, producing a local plasmon effect that enhances tumour-derived EV detection sensitivity and specificity ( Figure 2B). The unique plasmonic color of gold nanoparticles (AuNPs) and gold nanorods (AuNRs) allows the duplexed profiling of EV membrane marker, and their coupling caused red shift and color change can indicate the co-existence of the membrane antigens. They leveraged the Image J to analyze batches of DFM images. The method can automatically selected image areas with a brightness value of 255. After identifying these areas, the software calculated the ratios between the selected areas and the total areas of the images. These ratios were used as a metric for the specific nPES EV signal. In their study, the authors demonstrated that nanoplasmonenhanced scattering assay could effectively distinguish early-stage pancreatic cancer from pancreatitis and controls in a cohort of 155 subjects. However, the assay capacity for high-throughput surface marker profiling is limited considering the proximity of surface antigens and the steric hindrance of AuNPs and AuNRs during recognition and binding to the exosomal surface.
Gooding group 67 developed a core-satellite formation of Au nanoparticles (i.e., 67 nm-AuNP-analyte-enzyme coated 10 nm-AuNP) for the detection of interleukin 6 (IL-6). They utilized a CMOS-equipped digital camera and a darkfield microscope to analyze thousands of dropcasted-gold nanospheres within seconds using Image J ( Figure 2C). The research team developed a MATLAB script to transform images from the RGB color space into the HSV (i.e. Hue, Saturation, Value) color space, facilitating the separation of intensity and color variables. Following this conversion, a bandpass filter was applied to remove noisy pixels and enhance the image quality. The next step involved the identification of single nanoparticles based on their distinct morphology and intensity of the point-spread-functions. Each identified nanoparticle was Medicine Figure 1. Schematic illustration of confinement based ultrasensitive optical detection then labeled, and its color data (hue) is extracted using an intensity-weighted averaging method. This process entails creating a mask around the pixel with the highest intensity within a 3-pixel radius. By employing this robust process, the research team can independently count nanoparticles of various colors, enabling a comprehensive assessment of the nanoparticle population in the image. Compared to the LoD of ELISA and Western blot (~0.5 ng/mL), this dark-field imaging approach shows a 50-fold decrease in LoD due to the plasmon coupling. This method can circumvent the influence of non-specific adsorption due to the localized surface plasmon resonance (LSPR) sandwich assay, which is tolerant to non-specific binding effects, a major problem in the field of biosensing.
Wang group 68 designed another plasmonic nanobiosensor based on Au@Ag core-shell structure for detecting microRNA 21 (miR-21) ( Figure 2D). After capturing miR-21, an average localized surface plasmon resonance (LSPR) scattering wavelength shift of about 0.4 nm can be obtained. In this system, the "pyramidal" DNA tetrahedral structure was utilized as the recognition probe for miR-21 capped on noble-metal nanoparticles (e.g., gold or silver nanoparticles) by employing the shifts of the LSPR scattering spectrum peak (λ max ) as the output signal, permitting real-time detection of miR-21 with sensitivity at the aM level within a large dynamic range of 1 aM -1 nM. However, the DNA-microarray platform could only give rise to tiny LSPR scattering spectral wavelength shift (~0.4 nm) on this nanobiosensor, which may not be enough to be captured intuitively by a common instrument.
Plasmon coupled with fluorescence. Noble metal-based plasmonic particles can not only provide signal output by themselves, but also have the unique ability to couple with other fluorescent emission luminogens or quantum dots to enhance emission properties, permitting dark-filed and fluorescence dual-channel signal output. The physicochemical properties of plasmonic nanoparticles are specifically designed to facilitate plasmon-enhanced fluorescence (PEF) pathway, 49 which can significantly enhance the fluorescence of nearby fluorophore zone using spacers like SiO 2 or polymers with appropriate manipulation for the avoidance of quenching. While conventional fluorescent molecules and optical nanoparticles suffer from low brightness, poor photostability, and photobleaching, plasmon-enhanced fluorescence offers improved performance in terms of higher sensitivity and long-term photo-stability. Quantum dots (QDs) are a promising class of fluorescent probes for biological imaging owing to their unique stokes shift, powerful stability, and relatively long fluorescence lifetime. However, the fluorescence intensity of a single quantum dot is not sufficient for observation under a conventional optical microscope. Wang group 51 presented bioorthogonal nanoparticle detection (BOND) technology in a PEF-based suspension microarray using QDs as a signal reporter. The spherical gold-coated magnetic beads and flat gold film-coated glass were compared and served as solid support. They were decorated with anti-PSA (prostate-specific antigen) capture antibodies, followed by recognition and binding of the target analyte and the QDs-modified detection antibody, generating the final sandwich immunocomplex. After collection of fluorescence images of the plasmonic slide, the imagines were analyzed by software Genepix 6.1. This software can automatically identify the features and report the fluorescence intensities of each spot with background correction. After collecting the intensity data for each point, the mean fluorescence intensity was calculated for each set of spots, which can be converted to original concentration using Poisson equation. The designed platform utilizes BOND-and PEF-amplification strategies for ultrasensitive detection of proteins with detection sensitivity (1 fM) and dynamic range (5 orders of magnitude) was achieved for PSA detection ( Figure 3A). They demonstrated that both spherical and flat plasmonic surfaces can enhance the fluorescence of QDs, and the immunocomplex served as an excellent spacer.
To ensure multiplexed profiling, Shao group 62 developed a templated plasmonics for exosomes (TPEX) platform for multiparametric exosome analysis. This advanced method has been employed to analyze samples from 20 subjects (12 with colorectal cancer and 8 with gastric cancer), and has demonstrated its potential in predicting patient prognosis using exosomal proteins (MUC1, EpCAM, CD24, CD63) ( Figure 3B). Exosomes were incubated with fluorescent molecular probes and AuNPs. The exosome-bound AuNP can develop into a gold nanoshell-coated exosomes by seeding method, and the authors demonstrated that a large red shift in its plasmonic resonance can effectively quench the fluorescence signal of probes attached to the same vesicle. Besides, all the experimental steps were integrated into the miniaturized microfluidic chip that facilitated TPEX measurements of target in the complex clinical biofluids. A custom-designed smartphone-based optical detector can then be used for the image-based data acquisition and analysis, which can be completed within 15 min.
Recently, Singamaneni and co-authors 52 showed that a plasmonic nanoscale construct based on bovine serum albumin (BSA)-fluorophore decorated AuNRs which served as an "add-on" label for a broad range of bioassays with improved signal-to-noise ratio and dynamic range without altering their workflow and readout devices ( Figure 3C). In this work, BSA was used as a scaffold in the design of nanoconstructs, acting as a stabilizing agent to prevent nanoconstruct aggregation and as a blocking agent to minimize non-specific binding of the plasmonic-fluor. Compared to ELISA, the presented plasmonic assay exhibited 189-fold lower LOD and more than two orders of magnitude larger dynamic range. Such device was successfully utilized to scan an array of biomarkers associated with kidney disease. However, it should be noted that the exposed fluorophore on the AuNRs surface may be suffered from photostability, a protection shell would be favorable. Zheng and co-workers fabricated Ag@SiO 2 core-shell structure dopped with FITC, which are single ultrabright green particles that can be visualized by a conventional fluorescence microscope. This system permits long-term stability (6 months) with negligible attenuation of fluorescence intensity due to the silica shell protection and the resistance to the dye desorption. 63

Magnetic beads and nucleic acid amplification based detection
In addition to the plasmonic reporters, there are other non-plasmonic reporters available, which can be complementary candidates and have been actively explored for ultrasensitive assay. In the context of using fluorescence probes for ultrasensitive detection, traditional method using repeated labeling of fluorophores is often laborious and can lead to photobleaching, false-positive signal as well as a low signal-to-noise ratio 69 . Consequently, there is a pressing need to amplify the fluorescence signal toward singleentity detection in a robust way. In such regard, nucleic acid-labelled fluorophore-based detection probe in cooperation with solid supports like magnetic beads is a good option. Nucleic acid amplification (NAA) is a valuable technique for detecting pathogens or other biomolecules by amplifying specific nucleic acid sequences. 70 Numerous isothermal NAA techniques, including the catalyzed hairpin assembly (CHA), 71 hybridization chain reaction (HCR), 72 and rolling circle amplification (RCA), 73,74 have emerged as promising alternatives for achieving rapid and efficient signal amplification without the need of thermocycling. The Walt group 75 developed an ultrasensitive single-molecule detection platform that employed rolling circle isothermal amplification and a sandwich structure for Brachyury detection, achieving an impressive sensitivity of 244.6 aM ( Figure 4A). In this design, magnetic microbeads (MBs) conjugated with antibodies were used to probe human cytokines interleukin-1β (IL-1β) and IL-10. Then, single-stranded DNA (ssDNA) with a specific detection antibody and an RCA trigger were used to capture target exosomes. The RCA reaction enabled numerous fluorescent probes to hybridize with the RCA products, amplifying fluorescence signals. Beads were then placed on a slide and dried, forming monodispersed beads for digital counting using colocalized red (MBs)-green (probe) fluorescence. Following the similar principle, the Gao group 76 introduced modifications by using an aptamer as a recognition element instead of antibodies. They employed an addressable DNA nanoflower (DNF) attached to the surface of magnetic beads, which introduced a large number of fluorescence probes onto the nanoflower through base complementary pairing. This configuration enabled fluorescence imaging for the detection of 17β-estradiol (E2). a typical environmental estrogen, and a small molecule model target ( Figure 4B). The fluorescence image was automatically analyzed by Fiji image analysis software, simplifying the testing workflow. The surface confinement strategy, which involves using NAA and attaching optical probes to solid support, is promising but comes with a set of challenges. The intricate and prolonged wash procedure during binding and polymerization stages often results in significant product loss. Moreover, the inherent negative charge of DNA tends to non-specifically adsorb positively charged molecules, leading to undesirable false positives and high background interference. There is also a notable heterogeneity in the brightness of individual beads due to different amplification cycles, stemming from the sensitivity of enzyme activity to environmental influences.
To tackle these issues, appropriate antifouling strategies should be employed, such as the use of nonionic and non-protein blockers to suppress non-specific adsorption. [77][78][79] They can help reduce DNA-induced false positive signals, thereby pushing the detection limit further and improving the current fluorescence threshold. Automation for capture-detection would be another solution to overcome the multi-step manual handling caused errors. Additionally, rational fundamental design on sequences, the recently proposed one-pot amplification strategy might be a good alternative for its further evolution and improvement. 80,81 It is anticipated that RCA can be further improved by using signal enhancement and noise reduction approaches. Finally, the ability to control the efficiency of aggregation on individual magnetic beads results in discrepancies in fluorescence intensity among beads, which can have a substantial influence on the outcomes. It is worth noting that reactions such as RCA, LAMP, and PCR all require enzymes, whose activity can be easily affected by the environment, leading to low stability and assay accuracy. Hu et al. 82 ingeniously utilized size encoding approach to accomplish multiplexed detection by employing microbeads of four different sizes, followed by decoration with aptamers to target extracellular vesicle (EV) surface markers ( Figure 4C). They designed a microfluidic chip featuring an array of differently sized channels, which allowed magnetic beads of corresponding sizes to enter separate channels to reach their blockade with spatial resolution. This combination of size coding and spatial coding permitted simultaneous analysis of PD-L1, EpCAM, and CD63 from intact EVs for cancer diagnosis (healthy control n=10, cancer patients n=25). RCA amplification with hybridization of complementary DNA probe endowed the signal amplification capability. This study demonstrated that this tEV phenotyping method can rapidly and simultaneously detect six different tEV phenotypes with high sensitivity (Table 1).

Visible microsphere
In the aforementioned discussion, signal amplification techniques are highlighted, which enables the conversion of undetectable signals into observable fluorescent outputs through confinement-based signal enhancement. This method involves directly doping reporters like fluorescent tags to microspheres. The number of doped fluorophores and their intrinsic quantum yield determines the brightness of resultant fluorescent microbeads. Alternatively, the microbeads can be used as reporters as they can easily be visualized by conventional microscope. Noji et al. 83 introduced a sensitive wash-free detection method for PSA ( Figure 5A). In their work, PSA reacted with antibodycoated magnetic nanoparticles (~800 nm) and were then pooled into a femtoliter-sized reactor by magnetic force. The images were analyzed using Image J and Trackmate softwares. These tools can localize the center of individual particles by calculating the intensity of light spots. As a result, the number of tethered particles could directly correlate with the concentration of the target antigen. The number of particles tethered was proportional to the concentration of the target antigen. This method directly treated magnetic microbeads as reporters without the need of additional dyes. However, this method requires extensive mathematical calculations of the Brownian motion of molecules and imaging validation through other instruments. Yiping Chen 84 and his team encoded multiple targets using polystyrene microspheres with five different particle sizes, and the unconjugated microspheres in the supernatant after immunoreaction were recorded for subsequent visible signals under an optical microscope ( Figure 5B). The obtained images were further analyzed using artificial intelligence computer vision technology to decode the inherent characteristics of the polystyrene microspheres and revealed the type and concentration of the target. This method sidesteps the complexities of chip production and signal amplification but requires precise size control of the microspheres. Non-specific adsorption and precipitation must also be considered. The bright-field microscopy is straightforward but impurities or dusts can interfere, offering opportunity for the fluorescently encoded microspheres. To this end, He et al. 85 directly labeled magnetic beads (~1 µm) with fluorescent microspheres (~120 nm) to capture the target using a components of capture DNA-bead, analyte, and detection DNA-fluorescent microsphere in an order ( Figure 5C). Using SARS-CoV-2 ssDNA as the model, they achieved a 1.5 fM detection limit by counting and decoding bead-microsphere pairs and analyzed the data with NIH ImageJ software. 86 This study highlights the potential of single-particle luminescence for single-molecule detection. Noteworthy materials include aggregation-induced emission (AIE)  87 are anti-Stokes luminescent materials that transform low-energy nearinfrared (NIR) radiation into higher-energy radiation. They are favorable for detection with reduced interference and resistance to photobleaching. Carbon dots (CDs) 88,89 are photoluminescent nanomaterials that are watersoluble, biocompatible, and photostable, fitting the need for biological and environmental sample detection. Ongoing research in luminescent materials will bolster single-particle detection, broadening its applications in diagnostics, environmental monitoring, and chemical sensing. To conclude, the external confinement strategies widely employ micro-/ nano-architectures with the adhesion of optical reporters (e.g., organic dyes, plasmonic nanoparticles, quantum dots, etc.) on to their outer surface. In cooperation with signal amplification strategies like nucleic acid amplification or surface plasmon-enhanced fluorescence, the signal output can be dramatically enhanced to permit the stable probing of target at ultra-low levels, even single-entity level. It is expected that more novel formats on the nanoarchitecture design or signal enhancement and the cooperation with internal confinement will be introduced to achieve a synergetic effect, which would be the next hot spot.

INTERNAL CONFINEMENT STRATEGY
Besides the external confinement strategy, the internal strategy has also been actively explored as a vital complementary way. In such system, the signal output reporters, e.g., fluorophores or catalyzed photonic/colorful product, are usually confined within a small minuscule internal compartment of an architecture like microwell, droplet, metal-organic-frame (MOF), hydrogel, etc. The signal output molecules are restricted, producing localized ultrahigh concentrations with a resultant amplified signal. In the case of dropletbased systems, soft materials like emulsions or hydrogels are used to confine the signal molecules, restricting the signal output within the liquid droplet boundaries. For microwell-based systems, rigid materials such as plastics or silicon are used to form an array of small wells, limiting the signal output within the confined space of individual wells.

Confinement within droplet
In droplet system, mineral oil wrapped aqueous phase is commonly used. By application of positive external pressure to the inlet or negative pressure to the outlet generated through syringe pump 90 or syringe with a binder clip 91 , respectively, the stable water-in-oil droplet format can be formed. Only if the enzyme and substrate are encapsulated in the same droplet, the enzyme can catalyze and produce a fluorescent product thus illuminate the droplet under excitation. Then, digital readout ("1" or "0") can be generated once the number of droplets is far more than that of the target analyte. 92 Zheng et al., 93 utilized the capture antibody (anti-CD63)-detection antibody (anti-GPC-1) colocalization strategy for the single-exosome detection with an achievable LOD down to ~10 −17 M and a dynamic range of 5 log of the linear regime ( Figure 6A). Although the droplet system using enzyme-catalyzed fluorescence within a confined volume can generate a strong signal for ultrasensitive measurement and save sample consumption, it requires complex manual operations, e.g., the stepwise capture processes must be done outside the chip, and it only permits singplexing using current format. In such regard, the capability for automation, integration, and multiplexed profiling is favorable. Moreover, to further enhance the effectiveness of the dropletbased chip, it is crucial to develop a matched mobile platform that facilitates fast and easy access to assay data by integrating mobile device, thus highlighting the importance of improved connectivity and user-friendly interfaces. The Walt team 94 has developed a platform for rapid, sub-picogram-permilliliter, multiplexed digital droplet detection by using color-encoded beads and droplet microfluidics comprising of partition-incubation-detection compartments ( Figure 6B). This innovative approach employs a mobile phone-based imaging technique and coding-based duplexed profiling, eliminating the need for costly benchtop optics and ensuring a detection speed 100 times faster than conventional assays due to time domain-encoded mobile phone imaging. The performance of this assay was characterized by spiking recombinant proteins in fetal bovine serum (FBS) with a limit of detection (LOD) at 4 fg/mL (i.e., 300 aM). By integrating the coding-based duplexed profiling, they further optimized the system performance, enhancing its detection accuracy and reliability. This development will aid its commercialization, making a user-friendly system for self-testing anytime and anywhere. Liang's group 95 devised a microfluidic system, known as digital droplet with auto-catalytic hairpin assembly (ddaCHA), for the digital quantification of single microRNAs (miRNA) ( Figure 6C). This process involves partitioning miRNA molecules into individual picoliter-sized droplets, generating a digital readout in which the target molecule is classified as either present ("positive") or absent ("negative"). The researchers successfully implemented an enzyme-free auto-catalytic hairpin assembly (aCHA) within the droplets, demonstrating its effectiveness in minimizing the impact of external environmental factors and temperature fluctuations on droplet performance. In their investigation, an automated Python program was employed to monitor and compute the fluorescence intensities of individual droplets when they are flowing through a high-throughput microfluidic chip. By assigning a distinct identifier to each droplet within the flow channel, the script ensured that each droplet was only quantified once, thereby eliminating the risk of replicated counting. The significance of automation in such detection systems is obvious as it considerably reduces manual labor and minimizes potential errors in the process.
However, the complexity of these systems can be daunting for untrained end users, prompting the development of new approaches that aim to simplify or automate the microfluidic steps. Wang group 96 presented a novel lab-on-a-particle assay that employs hydrogel particles for simultaneous immobilization of protein molecules and emulsification templating, enabling digital counting of β-galactosidase. Following this process, image analysis algorithms in MATLAB are used to analyze the fluorescence signal of each droplet by averaging signals over the particle. A signal is considered to be positive if it exceeds 2.5 × standard deviations above the mean of the background signals ( Figure 6D). Moving forward, the efforts will be paid for refining this process and measuring actual targets further using appropriate affin- ity capture probes.
Droplet-based single-molecule detection has shown great potential in various applications, however, it presents certain drawbacks. For instance, droplet generation and manipulation can be technically challenging, and droplet stability may be affected by evaporation or coalescence over time. 97,98 Additionally, droplet-based approaches may not be suitable for all sample types or analytes, such as those with high viscosity or complex molecular structures due to challenges in generating uniform droplets and potential loss of analyte integrity. Furthermore, recovering analytes from droplets for further analysis could be difficult. These limitations call for the need of alternative approaches like microwell-based single particle detection, which may provide a more robust and versatile solution for many applications.

Confinement within microwell
Microwell strategies serve as an alternative to droplet-based techniques by confining chromatic or luminescent products within a tiny-volume microwells rather than the water-in-oil droplets. The microwell system is usually generated by nano-/micro-fabrication. Microwell-based techniques offer a controlled and stable environment for target analysis, as they do not rely on droplet generator and the complex manipulation. By employing microwell arrays, researchers can effectively isolate and analyze individual particles. Microwell confinement-based detection was originally developed by the Walt group 99 . This detection method, also referred to as Single Molecular Array (Simoa), has been successfully commercialized. In the Simoa approach, each femtoliter-sized reaction chamber in the array contains a single magnetic bead, and the number of microbeads (MBs) is deliberately maintained to significantly exceed the number of target molecules from highly diluted sample. In essence, it utilizes the concept of transformation of on-bead confinement to the micro-well confinement system, the enzymatically catalyzed fluorescence can be monitored by ultrasensitive CCD camera. The beads with the sandwich complexes are then loaded into these chambers, with each chamber designed to accommodate one bead. Subsequently, a fluorogenic substrate is introduced, which allows the enzyme on the detection antibody to convert the substrate into a fluorescent product. This process results in a binary "0" or "1" signal output, corresponding to the absence or presence of the target analyte, enabling highly sensitive detection. However, the Simoa system highly relies on the pricey bulk instrument and micro-well plates, which is not affordable to many remote areas and sourcelimited regions. Also, the current format is mainly designed for singleplexed profiling while multiplexing ability is still limited.
As microwell detection has evolved with continuous improvements and refinements, researchers have explored its potential in various applications. For instance, Karen group 100 has developed a novel digital microfluidics (DMF) chip-based method for detecting thyroid-stimulating hormone (TSH), a critical marker for assessing thyroid function. Such device requires a tiny sample volume of 1.1 µL and offers a LOD at 1.3 nanoIU/mL ( Figure 7A). Microwells were fabricated on a PET substrate using a UV imprinting process and resist material through a roll-to-roll fabrication method. In the DMF chip, precise control of electric fields between the top and bottom electrodes enables efficient droplet manipulation within these microwells, and the hydrophobic-inhydrophobic microwells on the chip also boost a higher seeding efficiency (97.6% ± 0.6%) due to minimized surface energy, ensuring stable droplet confinement, reduced merging or spreading. In another study, the Sun group 101 has developed a microfluidic chip using polydimethylsiloxane (PDMS) based microwell patterned arrays to detect tumor necrosis factor α (TNF-α) ( Figure 7B). Beads were injected into a chip and were allowed to settle for several minutes due to gravity. Subsequently, oil was injected through the inlet to seal an aqueous solution containing a single bead in the well. This power-free oil-sealed water phase eliminates the need for external pumping systems or valves, making it a cost-effective approach. Furthermore, the fabrication cost of these chips is low, and there is no need for complex surface modifications. The method also maintains good sensitivity with a limit of detection (LOD) of 12.62 fg/mL, which is comparable to performance using the commercial Simoa HD-1 Analyzer (14 fg/mL). In order to amplify optical signals and increase sensitivity, Li group 102 used edgeenhanced microwell (40 μm depth) immunoassay for detection of interferonγ ( Figure 7C). The microwells are coated with capture antibodies, which facilitate the binding of IFN-γ and the following Texas conjugated detection antibodies to form sandwich immune-complex. Due to larger axial resolution along with the vertical detection, the fluorescence emission was enhaced, then the signals of the analytes will be amplified greatly, which formed fluorescent rings. The edge enhancement effect of the microwell vertical sidewall shows a 6-fold sensitivity enhancement compared to those obtained on a flat surface.
It is worth mentioning that all the examples presented above are focusing on detecting single target, which may not reflect disease evolution and progression accurately. Also, lots of diseases originate by complex multiple factors, thus there is a need for multiplexed assays to permit multi-analyte profiling. In conventional assays, the detection of multiple biomarkers from one set of samples is challenging due to the complexity of encoding and decoding of numerous signal output from different analytes, and various error sources can lead to a decrease in the accuracy of the assay. 54 Therefore, rational design that allows distinguishable signal readouts, decent sensitivity, and specificity is the key pathway. Recognizing the existing limitations, Sung group 103 introduced the pre-equilibrium digital ELISA (PEdELISA) microarray platform that leverages spatial-spectral encoding and machine learning on a microfluidic chip ( Figure 7D). This allows 12-cytokine assay using only 15 μL of serum. The 12-plex PEdELISA microarray uses a CNN-based parallel computing algorithm, ensuring unsupervised image data analysis with high accuracy. It autonomously classifies and segments image features at a high throughput (1 min/analyte), achieving 8-10 times higher accuracy than traditional GTS-based algorithms without manual corrections. The entire detection and analysis cycle can be completed within 40 min. In essence, PEdELISA enhances multiplexing efficiency by integrating advanced microfluidics with a machine learning image processing method.

AI APPROACH FOR OPTICAL DETECTION IN CLINICAL ANALYSIS
Indeed, with the remarkable advancements in hardware technology, there has been an extraordinary surge in the development and implementation of artificial intelligence (AI) tools that are specifically designed for disease diagnosis. These bespoke AI systems excel at processing vast volumes of data and images sourced from a myriad of equipment. By incorporating AI models, such as Linear Discriminant Analysis (LDA) [104][105][106] , Support Vector Machines (SVM) 107,108 , Logistic Regression, Linear Regression, Linear Mixed Models and Random Forest Models (RFM) 109,110 , these systems significantly enhance both the efficiency and accuracy of diagnostic procedures. Within the scope of our specialized system, the primary objective is the automated detection and quantification of visible particle signals as particle count can act as a surrogate for biomarker concentration. By accurately quantifying specific protein biomarkers in patients, these AI-driven approaches can significantly advance medical research and patient care, potentially leading to earlier interventions and improved treatment outcomes. The schematic diagram of this AI application in our review can be summarized as follows: Image acquisition → Particle detection and counting → Biomarker concentration representation → Data Analysis→Diagnosis (Figure 8). This progression succinctly captures the role of AI in augmenting diagnostic procedures in healthcare.

Automation and standardization of image processing
AI technology has revolutionized optical detection by automating data processing, boosting efficiency, and reducing human error. In image process-Medicine Figure 8. The general process of AI techniques for data interpretation REVIEW ing, AI excels in automating particle counting through several steps: (1) Image Preprocessing: AI algorithms enhance image quality by decreasing noise and elevating contrast. Techniques like Gaussian filtering, thresholding, and contrast normalization 111,112 are used for clarity, facilitating better detection and accurate particle counting. (2) Object Detection: Depending on image complexity, methods range from connectivity analysis for simple images to advanced techniques like edge detection and machine learning for intricate ones. (3) Object Counting: After segmentation, objects are enumerated using tools like ImageJ/FIJI, MATLAB, and Python libraries such as OpenCV and Scikit-Image. The Chen group 113 demonstrated this, differentiating polystyrene (PS) microspheres with a CV technique, achieving 93.5% to 100% accuracy, surpassing traditional methods. For example, the Chen group 113 utilized computer vision (CV) methods to differentiate and quantify polystyrene (PS) microspheres of varying sizes. ( Figure 9A) Through techniques like binarization thresholding and a customized algorithm, they achieved impressive accuracy ranging from 93.5% to 100%, surpassing traditional manual methods.
To sum up, ultra-sensitive optical detection can be realized through various rational signal amplification strategies, translating minuscule signals into countable particles, bypassing the need for complex AI technologies typically required in other systems. However, when a scenario necessitates more intricate analysis, such as distinguishing signals based on their intensity or correlating signal strength with specific biomarker concentrations, advanced algorithms and strategies become invaluable. Nevertheless, the incorporation of AI, even within these simplified environments and small-scale clinical sample analysis, significantly enhances the accuracy and reliability of our methodologies.

AI approaches to data interpretation
Analysis of collected biomarker data plays a crucial role in gaining a comprehensive understanding and diagnosis of diseases. The abundance and diversity of biomarkers offer valuable insights for disease diagnosis, monitoring treatment response, tracking disease progression, and identifying disease subtypes. With the advent of AI techniques, the analysis of biomarker data has become more efficient and accurate, enabling the delivery of personalized healthcare services to patients. In the following section, we will elaborate how various AI techniques work on these critical medical areas.
In the domain of early disease diagnosis, SVM-based methods are notably adept at dealing with high-dimensional data, like gene expression profiles. Their proficiency in extracting complex patterns from this data allows them to discern distinct gene expression signatures that could separate cancer cells apart from healthy cells. This distinguishing capability is pivotal in facilitating the early detection of cancer.
Given the prominence of high-dimensional gene expression data in cancer research, the application of SVMs might be more prevalent in this field than in others. 114,115 However, their performance depends on the correct selection of a kernel function and adjustment of parameters, which may require specialized domain knowledge. Further research in this area could greatly enhance the applicability of SVMs.
Moving on to the monitoring of treatment responses, Random Forest models play a central role in this domain. [116][117][118][119] These models are ensemble learning methods that combine multiple decision trees. This approach could provide an understanding of the intricate interactions between various biomarkers and treatment outcomes. For instance, when managing chronic diseases like diabetes, Random Forest models can effectively handle diverse data, ranging from glucose levels and insulin doses to lifestyle factors like diet and physical activity. 120 This capability could help develop a more personalized and effective treatment strategies. However, the complexity of the model can increase as the number of features or variables grows, making it challenging to interpret the results and understand the underlying patterns effectively. Therefore, future research may explore more effective ways to optimize the number of decision trees in the ensemble or propose new data dimensionality reduction techniques.
Linear Discriminant Analysis (LDA) plays an important role in tracking disease progression. [121][122][123][124] Its capability to simplify high-dimensional biomarker data makes it instrumental in identifying longitudinal changes in biomarker levels, paving the way for a deeper understanding of disease trajectories over time. For instance, when monitoring heart disease progression, LDA can be applied to unravel dynamic alterations in multiple biomarkers, such as cholesterol levels, blood pressure readings, and inflammation markers. This analytical power aids in understanding the fluid nature of disease progression and enabling timely interventions. For example, Jin group 125 employed a combined approach of principal component analysis (PCA) and Linear Discriminant Analysis (LDA) to extract features from highdimensional Surface-Enhanced Raman Scattering (SERS) spectra ( Figure  9B). First, they performed dimensionality reduction on the SERS spectra using PCA, reducing them to 46 principal components that accounted for 95% of the total variance. Next, they applied LDA to further reduce the dimensionality of these 46 principal components to just two LDA factors. Remarkably, this combined approach resulted in a clear separation and accurate identification of exosomes derived from normal cells and cancerous cells.
In the field of disease subtyping, linear and logistic regression models serve as potent tools for grouping. 126,127 These models can shed light on the associations between biomarkers and specific disease subtypes. In addition, these models can integrate an array of biomarkers, such as gene expression profiles and metabolic markers, to estimate disease subtypes, enabling more effective and personalized management plans.
In conclusion, AI techniques have demonstrated their effectiveness in analyzing biomarker data, facilitating a deeper understanding and diagnosis of diseases. These techniques contribute to early disease detection, treatment response monitoring, disease subtyping, ultimately providing more accurate and tailored healthcare services for patients. However, it is also essential to be aware of potential challenges. AI models heavily rely on the quality of the training data. Therefore, ensuring the accuracy, completeness, and standardization of biomarker data is critical to yield reliable results. In addition, these models often involve complex mathematical computations and require domain knowledge for optimal parameter tuning, highlighting the need for interdisciplinary collaborations between clinicians, data scientists, and bioinformaticians. Another potential challenge in embracing AI advancements is overfitting due to sample size. Overfitting occurs when a model is excessively complex and learns noise or specific patterns from the training data that do not generalize well to real data.
Another area that warrants attention is the interpretability of these models. While AI models can make highly accurate predictions, the underlying reasoning can be opaque, a phenomenon often termed as the "black box" problem. This lack of transparency can hinder physicians' trust in the model predictions, thereby limiting their adoption in clinical practice. Lastly, ethical and legal considerations surrounding data privacy and security, and algorithmic bias should also be considered when deploying AI in healthcare. As we move towards this future, it is imperative to address these challenges headon to fully harness the potential of AI in transforming patient care and outcomes. The amalgamation of AI with medical science is truly an exciting frontier. It promises not only advancements in disease understanding and management but also a new era of personalized medicine. With careful navigation, we stand on the cusp of a healthcare revolution.

CONCLUSION AND PERSPECTIVE
This review provides a comprehensive view of confinement-based ultrasensitive optical assays (Table 1), enabling the (digital) measurement of analytes without relying on advanced imaging systems such as total internal reflection fluorescence (TIRF) 43,44 microscopy or photo-activated localization microscopy (PALM). 128,129 As the investigation of novel and uncommon disease markers expands, along with the need to decipher multifactorial diseases, there is a rising demand for the development of unsophisticated, ultrasensitive, and multiplexed detection platforms to address these complex challenges. Luminescent systems, either through self-illuminating compounds or through the products of enzymatically induced luminescence or color systems, can be confined to internal or external spaces. These systems can cooperate with numerous signal amplification approaches, such as incorporating surface plasmon enhancement and nucleic acid amplification, to realize ultra-bright and ultra-strong signal readouts. The core aim is to harmonize the cooperation of signal amplification with noise reduction, reaching a balance that is crucial to achieving a distinguishable signal from the background at a single-entity level. This balance plays a key role in enabling early diagnosis and treatment. However, this cutting-edge detection approach still faces significant obstacles and challenges, highlighting the need for further investigation and resolution to fully exploit its potential in diagnostic and therapeutic applications.
Ultrasensitive optical detection requires specific probes to avoid false results. Enhancing probe specificity can be achieved through unique molecular design, directed evolution, high-throughput screening, multi-domain probes, and computational modeling. These steps can substantially improve the accuracy and reliability in applications like disease diagnosis. Various signal amplification strategies aim for ultrasensitivity, with the reduction of background noise being paramount. Incorporating anti-fouling techniques, such as PEGylation and zwitterionic coatings, can minimize contaminants, enhancing detection accuracy. 130 With the rise of automation and AI, their synergy with ultrasensitive detection is gaining traction. As the detection process requires sample purification to eliminate interference, automation offers improved efficiency and accuracy over traditional methods. AI is particularly adept at rapidly interpreting images obtained through microscopy. Moreover, training AI on vast patient datasets allows for predicting disease trajectories and crafting tailored treatment plans. By fusing AI with particle-guided optical assays, we pave the way for innovative disease diagnostics and treatments. While many of these technologies remain in their nascent stages, leveraging larger sample sizes in the future could lead to significant breakthroughs. This evolution symbolizes the shift from mere big data to its profound mining, optimizing the union of AI and detection hardware. 131 The intersection of biotechnology (BT) and information technology (IT) is undeniably shaping our present and future. To achieve impactful outcomes, strategies in confinement, signal amplification, and encoding must seamlessly integrate. The overarching goal is to craft miniaturized, automated, and smart systems for monitoring evolving disease threats with validated markers, a venture that underscores the importance of interdisciplinary teamwork. The next decade promises transformative advancements, especially in singleentity detection technology. 132 With potential applications spanning molecular biology, biomedicine, to nanotechnology, it is essential to address any challenges and champion continued innovation for its broader acceptance. Such progression is vital for proactive disease management, improving human health, and realizing novel technological leaps.