The reagents used in various experiments in this research are detailed here. Bovine serum albumin (BSA) was obtained from Sangon Biotech (Shanghai) Co., Ltd. PD-L1 aptamer (H0) and HI and H2 (both simple and biotin labeled) sequences obtained from reference (43) were ordered from Sangon Biotech (Shanghai) Co., Ltd and are tabulated in Supplementary Table S1 . Other auxiliary reagents such as HEPES buffer and SuperBlock blocking buffer (cat log no: 37581) were ordered from ThermoFisher Scientific, and Streptavidin-labeled horseradish peroxidase (HRP) was from Bioss Antibodies. Tetraethoxysilane (TEOS) and tetrabutyltitanate (TBOT) were purchased from Sinopharm Chemical Reagent Co., Ltd (China). Dopamine was purchased from Sigma-Aldrich. Tween-20 was acquired from Beyotime. Salmon sperm DNA was ordered from Invitrogen. For the preparation of various buffers used in this study, we used pure water obtained through the Milli-Q benchtop water purification system. All chemicals were of analytical grade and were used as received without further purification.

A549 (human lung adenocarcinoma cells), BEAS-2B (human lung epithelial cells), and H596 cells were purchased from Shanghai Cell Bank, Chinese Academy of Sciences. To culture the cells, Gibco’s Dulbecco’s modified Eagle’s medium (DMEM) was used. Fetal bovine serum (FBS) and exosome-depleted FBS were delivered by Gibco and System Bioscience, respectively. The culturing of the cells (A549, H596, and BEAS-2B) was achieved at 37°C in 5% CO₂ and 95% air using DMEM containing 10% FBS (44). Once the growing cells covered almost 85% of the culture plate surfaces, we discarded the used media, washed the plates thrice with PBS, and added fresh media (DMEM) supplemented with 10% exosome-depleted FBS. After the 24- to 48-h incubation time, the media was poured into 50-mL conical tubes, stored at −20°C, and later subjected to ultracentrifugation to collect EVs (45).

Ultracentrifugation-based isolation of EVs from media involves spinning of culture media successively at 300 × g for 10 min, then 2,000 × g for 10 min, and 10,000 × g for 30 min (in Beckman Coulter, Allegra X-30R) at 4°C to mitigate unwanted components such as cells, dead cells, and cell debris, respectively. Following these steps, the clarified supernatant was filtered using a 0.2-µm syringe filter. This filtered solution was then ultracentrifuged (100,000 × g for 90 min) at 4°C using a Beckman Coulter Optima XPN-100 centrifuge. The supernatant was removed, while the remaining pellet, enriched in EVs, was resuspended in 1 mL of PBS and subjected to another round of ultracentrifugation (100,000 × g for 90 min) at 4°C using a TYPE 90 Ti Rotor. Finally, the supernatant was removed and the precious EV pellet was resuspended in 100 µL of sterile-filtered (0.22 µm filter) PBS before being stored at −20°C for experiments to be conducted immediately and at −80°C for experiments to be performed later.

The (Fe₃O₄₎-SiO₂-TiO₂ mag-nanoparticles were prepared following a combination of previously published protocols (19, 46). We prepared (Fe₃O₄)-SiO₂-TiO₂ mag-nanoparticles using Li et al.’s (24) protocol, including a classic recipe: the Stöber method. First, we took 0.3 g of Fe₃O₄ and mixed it up with 0.5 L of ethanol, 0.15 L of deionized water, and 0.01 L of ammonia. It was exposed to sonication for 15 min. Then, we added 4 mL of TEOS and continued to stir (1,500 rpm) the whole concoction for 10 h at 25°C.

Finally, (Fe₃O₄)-SiO₂-TiO₂ was made using a kinetics-controlled coating process. We started by taking 0.05 g of our (Fe₃O₄)-SiO₂ and mixed it with an amalgam of 0.1 L of ethanol and 0.3 mL of ammonia (28%). Similar to before, the mixture was sonicated for 15 min. Then, we added 0.75 mL of TBOT in ethanol while constantly stirring (2000 rpm, 24 h, 45°C) the mix. Finally, the (Fe₃O₄)-SiO₂-TiO₂ mag-nanoparticles were washed 3× with water and ethanol and then dried in an oven overnight at 70°C, making them ready for utilization in future experiments.

EVs’ shape and structure (morphology) were confirmed using transmission electron microscopy (TEM). Approximately 20 μL of the EV sample was carefully placed onto a copper grid. Next, we employed filter paper to gently remove any excess liquid. To enhance the contrast of the EVs in the image, a 2% phosphotungstic acid solution was dropped (5 min). Following that was a washing step using double-distilled water (5×) and then the copper grid was allowed to dry completely. Finally, the EVs were visualized using a high-powered JEM-2200CX TEM microscope from JEOL. For WB, protein concentration was assessed using a Qubit fluorometer from Invitrogen. Next, the EVs were subjected to a lysing step, by heating them in loading buffer at 95°C for 5 min. This process essentially breaks down the EVs and releases their protein cargo. The lysed proteins were then loaded onto a specialized gel (10% SDS-PAGE gel ordered from Sangon Biotech) specifically designed to separate proteins based on their size.

To capture these separated proteins, a polyvinylidene difluoride (PVDF) membrane, also ordered from Sangon Biotech, was employed to faithfully transfer the protein bands from the gel during a subsequent step. Subsequently, the membrane was blocked using a 2% non-protein blocking solution for 2 h. After blocking, the membrane was washed and dunked into a mixture of primary antibodies [CD9, TSG101 (ab275018 by Abcam), and PD-L1 by Proteintech] and kept at 4°C for 12 h. Excessive primary antibodies were washed (3×) away and the membrane was dunked in the solution of appropriate secondary antibodies (goat anti-rabbit, Abcam) at room temperature for 2 h.

Finally, the bands were visualized using the enhanced chemiluminescence (ECL) kit ordered from Sangon Biotech. To determine the EVs’ concentration and size distribution, we used an NTA machine (ZetaView, Particle Metrix, Diessen, Germany). EV samples were serially diluted (typically 1:100 to 1:1,000) using filtered PBS to ensure that the EV samples fell within the detectable range of the NTA machine. Subsequently, the ZetaView^® software automatically analyzed all the captured frames, ensuring that the outliers discarded the most relevant parameters; that is, the concentration of EVs and their average size were calculated.

Performed via Leica Microsystems confocal microscope equipped with a 100× oil immersion objective, EVs were dyed using PKH26 dye following the manufacturer’s instructions. PBS was used as the control. A Zetasizer Nano series (Nano-ZS) instrument from Malvern was engaged to evaluate the surface charge of (Fe₃O₄)-SiO₂-TiO₂ mag-nanoparticles (0.5 mg), A549 cell-released EVs (10⁸ EVs/mL), and the complex formed between the mag-nanoparticles and EVs (0.5 mg/mL).

EVs enriched from the A549 cell culture media were combined with (Fe₃O₄)-SiO₂-TiO₂ mag-nanoparticles for testing their EVs’ capturing strength, using a previously reported method based on the PKH26 dye (19). According to the method, we made two samples with the same number of EVs isolated from A549 cells. In both samples, EVs were dyed with PKH26 dye following the manufacturer’s instructions. A specific concentration of EVs (10⁸ EVs/mL) was used for this step. The fluorescence of these dyed EVs was then measured using a spectrophotometer. Next, the dyed EVs were mixed with (Fe₃O₄)-SiO₂-TiO₂ mag-nanoparticles. After separating the mag-nanoparticles and the attached EVs using a magnet, the fluorescence of the solution (supernatant) was measured. For accuracy, the experiment was repeated three times.

Now, to test that how well the (Fe₃O₄)-SiO₂-TiO₂ captured the EVs, we calculated the capture efficiency (%) using the following formula: capture efficiency (%) = (F2/F1) × 100. Here, F1 represents the initial fluorescence of the EVs solution before being enriched with the (Fe₃O₄)-SiO₂-TiO₂ mag-nanoparticles, and F2 represents the fluorescence after enrichment. The experiment was further expanded to determine the ideal concentration of (Fe₃O₄)-SiO₂-TiO₂ mag-nanoparticles and the optimal incubation time needed to achieve the best capture efficiency.

To visualize the HCR product, we performed agarose gel electrophoresis and PAGE (Vazyme kit). For agarose gel electrophoresis, 2% agarose was used. H1 and H2 were separately pre-heated (at 95°C) for a short time and then left to cool at room temperature for an hour. For HCR, H0 (0.6 µM), H1 (2 µM), and H2 (2 µM) were mixed. Ten microliters was taken from each sample mixed with 2 µL of loading buffer. To load onto the gel, we took 10 µL from each sample and performed electrophoresis. Likewise for PAGE, 10 µL was taken from each sample and 2 µL of ×6 loading buffer was mixed. Later, 10 µL of the mixture was loaded into a 12% PAGE. Running was done at 100 V for 2 h in ×1 TBE buffer. After that, the gel was stained for 10 min in 50 mL ×1 TBE buffer with 5 μL of ×10,000 Gel-Red nucleic acid gel stain. The gels were then photographed using a Tanon 2500R gel imaging system.

We obtained ethical approval and written consent from the participants to use their serum samples, which were collected at the clinical laboratory of the Drum Tower Hospital affiliated with Nanjing Medical University. A total of 42 samples were acquired from 21 healthy and 21 lung cancer patient samples; their clinical details are tabulated in Supplementary Table S1 ( Supplementary Data Sheet 1 ). Initially, a 0.22-µm filter attached to a syringe was used to filter the serum. Then, approximately 50 µL of the unaltered serum was incubated with (Fe₃O₄)-SiO₂-TiO₂ for 10 min to capture EVs. Subsequently, the remaining steps of the designed colorimetric aptasensor were performed. We used PBS as a blank control for comparison.

For comparison purposes, we followed the conventional plate ELISA protocol of Chen et al. (10) for PD-L1@EVs, with slight modifications. PD-L1 antibody (5 µg/mL) was coated to the wells of the plate and kept at 4°C overnight. The SuperBlock solution was used for blocking for 2 h. EVs isolated from clinical samples were added to the wells and incubated overnight. The detection antibody was added at a concentration of 1 μg/mL per well and incubated for 1 h at room temperature. Afterward, 100 μL of HRP-streptavidin, diluted 1:10,000 in SuperBlock blocking buffer, was added to each well and incubated for 15 min at room temperature before measurement.

Deep learning is preferred for complex tasks as it automatically extracts information from raw data to learn and train and then utilizes those data to make decisions about new unexposed data (47). Therefore, a deep learning model was established for chromaticity analysis. We acquired images of tubes from PD-L1@ EVs concentrations ranging from 0, 10³, 10⁴, 10⁵, 10⁶, 10⁷, 10⁸, 10⁹, and 10¹⁰ EVs/mL. The images were obtained through three different types of smartphones (iPhone X, Huawei mate 30 pro and Redmi Note 12 Turbo) under standard lab lights using different shooting angels and default camera settings from 40 cm height. However, the background was consistently kept white. Moreover, through data augmentation, we increased the images to 3,000 for better training of the deep learning model using various conditions such as color temperature, brightness modulation, contrast, pixel, image sharpening, Gaussian noise, and motion blur. Moreover, we consistently used 1.5 mL of Flex-Tube by Eppendorf throughout these experiments.

The deep learning model needs the optimization of the data format and parameters (48). It was trained using the TensorFlow and Keras frameworks (49, 50). The images were converted into NumPy arrays suitable for TensorFlow processing, and later converted from an RGB to an HSV format. Color thresholding was adjusted to define a broader hue range ([5, 40]) to capture light brownish to blackish colors. Saturation ([30, 255]) remains high to ensure a strong color presence. The value ([20, 255]) allows for variations in brightness within the color range. Focusing on the colored region (our sample) was achieved using the find_largest_contour function and the center of gravity. Based on the center of gravity, the images were first cropped to 600 × 600 pixels and then resized to 128 × 128 pixels. The pixels were normalized between 0 and 1 and saved as.pkl file. Then, we randomly divided the data between the training set (80%, 2,400 images) and the testing set (20%, 600 images).

CNN typically has convolution and pooling layers. The function of the convolution layer is to extract features from the images (51, 52). Then, it has a maxpooling layer to select the maximum value within a receptive field (53). After that, fully connected layers combine the features obtained by the previous two layers. Fully connected layers consist of a denser connection of neurons that are connected with neurons in the following layer (54). A CNN achieves nonlinearity by the application of an activation function. In general, CNN use rectified linear units (ReLU), but here we used a scaled exponential linear unit (SELU) activation function. SELU is a self-normalizing activation function. It is preferred for faster learning and prevention of gradient vanishing (55). The model configuration is tabulated in Supplementary Table S2 .

The next step after training the model was the development of an app. For this purpose, we used our efficiently trained CNN model and TensorFlow Lite converter API to convert our Keras model into a TensorFlow Lite model (.tflite file), making it suitable for mobile phone devices. An app was developed using Android studio and Java. To integrate TensorFlow Lite, we used the TensorFlow Lite library for Android. For images obtained through the app to match the image size used for model development, we also preformed image preprocessing operation on the images imported through the app using the OpenCV_for_andriod library. Java was used to write images cropping from 600 × 600 to 128 × 128 pixels, as that of the image dimensions used for CNN training. BitmapFactory in Java programming language Android SDK (software development kit) was used to mitigate the impact on the image pixel quality. Finally, through the model’s prediction, the app’s user interface displays the estimated concentration of PD-L1@EVs.

The main design of the colorimetric aptasensor of EVs coupled with an intelligent deep-learning-powered smartphone app for quantitative analysis of chromaticity is depicted in Figure 1 . As depicted in Figure 1A , EVs were magneto-captured by (Fe₃O₄)-SiO₂-TiO₂ mag-nanoparticles. After blocking with SuperBlock for 60 min, the PD-L1 aptamer that binds to EVs with the PD-L1 marker was added. Once the PD-L1 aptamer is attached, the addition of H1 and H2 (biotinylated) initiates the HCR reaction (PD-L1 apt + H1 + H2), forming a long self-assembled DNA concatemer. The numerous HRPs (streptavidin-conjugated) attached to the long strand then produce a colored signal upon the addition of the HRP substrate. HRP uses dopamine (colorless) as a substrate and converts it into polydopamine in the presence of H₂O₂ (a 300-fold enhanced reaction in a short time) (42). The change in chromaticity (absorbance at 400) corresponds to the quantity of PD-L1@EVs. To transform the qualitative colorimetric approach into a quantitative operation, we developed an intelligent deep-learning-powered quantitative analyzer for chromaticity in the form of a smartphone app named ExoP, thereby achieving the intelligent analysis of chromaticity with minimal user intervention or additional hardware attachments for the sensitive and specific quantification of PD-L1@EVs ( Figure 1B ). We envision that our smartphone app ExoP can fill in the gap between quantitative and qualitative measurements. Using a smartphone app as a portable quantitative analyzer of chromaticity is helpful to overcome the limitations of the colorimetric technique and improve quantitative analysis. From successful initial investigations, it can be postulated that the proposed methods hold promise for the rapid and accurate analysis of PD-L1-EVs in clinical serum samples, potentially paving the way for improved lung cancer detection.

The first important element of the colorimetric aptasensor is the development of (Fe₃O₄)-SiO₂-TiO₂ mag-nanoparticles. To analyze the morphology and composition of our (Fe₃O₄)-SiO₂-TiO₂ mag-nanoparticles, we employed the powerful duo of SEM and EDS. SEM generates a detailed image of the nanoparticle surface, revealing their size, shape, and overall morphology. EDS, on the other hand, reveals the elemental makeup of the nanoparticles, identifying the specific types of atoms that are present. Supplementary Figure S1A shows the images taken with SEM and revealed that the (Fe₃O₄)-SiO₂-TiO₂ nanoparticles were uniformly round with a bumpy surface made of many tiny clumps of TiO₂. Furthermore, EDS ( Supplementary Figure S1B ) confirmed the formation of (Fe₃O₄)-SiO₂-TiO₂. The silica layer acted as a bridge, allowing for the later addition of TiO₂.

The XRD report shown in Supplementary Figure S1C suggests that the particles are crystalline and have the anatase phase. Peaks at 2 theta 30.88, 35.22, 43.06, 53.5, 57.1, 62.72, and 74 correspond to the planes of the cubic phase of Fe₃O₄ and anatase phase of TiO₂ (JCPDS No. 21-1272). The diffraction peak of (Fe₃O₄)-SiO₂ showed no significant change compared with Fe₃O₄ because the coating layer SiO₂ also belongs to an amorphous body. The XRD results suggest the successful formation of (Fe₃O₄)-SiO₂-TiO₂. FTIR-based examination of (Fe₃O₄)-SiO₂-TiO₂ nanoparticles ( Supplementary Figure S1D ) showed peaks at 802 and 1,068 cm⁻¹, which confirm the presence of silica coating on the magnetic nanoparticles. After the addition of TiO₂ to the surface, a new peak appeared in the FTIR spectrum at 970 cm⁻¹. This new peak indicates vibrations between the titanium and oxygen atoms, proving the presence of the TiO₂ layer. Peaks in the range of 500–700 cm⁻¹ can be attributed to Ti–O–Ti and Ti–O–Si vibration.

It has been recommended by MISEV 2023 guidelines to provide detailed profiling of EVs and their concentration. EVs acquired from A549 cells were assessed as cancer-specific EVs. While EVs obtained from the Beas-2B cells were assessed as normal control. TEM is one of the most widely used technologies for assessing the morphology of EVs. TEM is hailed for its prowess in detecting EVs, irrespective of size. The images obtained via TEM confirmed the presence of vesicular structures ( Figure 2A ). Another important aspect of EVs is to confirm the presence of EV-specific markers and cancer-specific markers. WB has been widely recommended for the detection of proteins in EV-containing preparations. It can be seen in Figure 2B that markers specific to EVs such as CD9 and TSG 101 are present. Likewise, the negative marker calnexin can also be observed to be present on cells but missing on EVs. We also observed the presence of PD-L1 markers both on lung cancer cell lysate (A549) and the EVs derived from those cells. NTA is an optical technique that provides information about EVs’ size and abundance. Figures 2C, D show the NTA analysis of the concentration and size of EVs obtained from BEAS-2B (10⁸ particles/mL) and A549 cells (10¹¹ particles/mL), respectively. In Figure 2E , EVs can be seen visualized through confocal microscopy using a membrane binding specific dye, PKH26. From the above panels, it can be concluded that we obtained both BEAS-2B EVs and A549-derived EVs at high concentrations for conducting further experiments.

One of the cornerstones of colorimetric aptasensor design is (Fe₃O₄)-SiO₂-TiO₂-mediated EV capture. Hence, we used the zetapotential techniques to confirm this. A significant change in the potential of mag-nanoparticles can be observed after the introduction of EVs, providing evidence of mag-nanoparticles binding the EVs ( Figure 3A ). Next, we systematically investigated the optimum quantity of (Fe₃O₄)-SiO₂-TiO₂ and the incubation time required for optimal capture efficiency. Varying concentrations of (Fe₃O4)-SiO₂-TiO₂ (0.2–1.6 mg) were employed to determine the amount necessary for maximum capturing of PKH26-labeled EV. The fluorescence intensity was measured before and after the addition of (Fe₃O₄)-SiO₂-TiO₂. As depicted in Figure 3B , a dosage of 1 mg yielded an optimal capture efficiency exceeding 90%. Further increments in bead concentration resulted in negligible improvements.

Incubation time of the (Fe₃O₄)-SiO₂-TiO₂ for peak capture efficiency was optimized using PKH26-labeled EVs. Incubation durations ranging from 2 to 20 min were evaluated. The maximum capture efficiency was attained at 10 min, with negligible gains observed beyond this point ( Figure 3C ). In addition, the temperature for best capture efficiency of (Fe₃O₄)-SiO₂-TiO₂ was optimized using various temperature conditions (4°C, 26°C, and 37°C). The mag-nanoparticles exhibited the best capturing efficiency at 25°C ( Figure 3D ).

To visually confirm the HCR, we performed Agarose gel electrophoresis ( Figure 4A ) and PAGE ( Figure 4B ). Both experiments provided visual proof of the HCR.

The PD-L1 aptamer (H0) has a central role for the successful initiation of HCR. A series of aptamer concentrations (H0 concentration from 0.2 to 1 µM) were evaluated against a constant concentration of EVs (10⁹ EVs/mL) as the experimental sample, with PBS acting as a blank control. The optimal signal-to-noise ratio was achieved at 0.6 µM concentration ( Figure 4C ). Subsequently, the optimal incubation time for the HCR reaction was optimized to maximize the S/N ratio. Three durations (30, 45, and 60 min) were tested. As illustrated in Figure 4D , the highest S/N ratio was attained at 45 min, with progressively higher background signals observed for longer HCR times.

After optimization, we explored the sensor’s analytical parameters such as sensitivity, specificity, and accuracy for detecting PD-L1@EVs.

A broad range of A549 EVs concentrations (10³–10¹⁰ EVs/mL) were used for analysis. The peak absorbance at 400 nm, indicative of color intensity, was precisely calibrated using a spectrophotometer. A change in color and the corresponding absorbance (at 400 nm) were recorded for various concentrations of A549-EVs as well as for a blank sample containing PBS without EVs as in Figures 5A, B .

It can also be seen in Figure 5C that the change in color intensity/absorbance and the concentration of PD-L1@EVs are related in a linear fashion ranging from 10³ to 10⁸ EVs/mL following the equation Y = 0.3167*Log N + 0.4031, where Y stands for peak absorbance and N stands for the number of EVs. Similarly, the LOD was estimated by incorporating the typical three times the standard deviation of blank samples in the above equation. The LOD was estimated to be 3.6×10² EVs/mL. It is of paramount importance for the sensor to be highly specific for target detection. We checked the sensor’s specificity by using EVs obtained from a variety of sources, such as EVs from the serum of a healthy volunteer (HS EVs), EVs from BEAS-2B cells (PD-L1-negative EVs), A549 cells, and H596 cells (two PD-L1-positive EVs). As can be seen in Figure 5D , high absorbance was recorded for the PD-L1@ EVs as compared to BEAS-2B EVs and FBS-derived EVs. Thus, it can be inferred that the designed colorimetric aptasensor demonstrated high specificity for the detection of PD-L1@EVs.

A critical aspect is to test the performance of the designed colorimetric aptasensor using complex clinical samples, such as serum. We tested the colorimetric aptasensor using a total of 42 clinical serum samples, of which 21 samples belonged to the healthy individuals and 21 belonged to the lung cancer patient group. Detailed clinical information is provided in Supplementary Table S1 (Supplementary Materials). In Figure 6A , the absorbance values were significantly lower (p< 0.001, Mann–Whitney U test) for the serum samples belonging to the healthy group compared to the lung absorbance recorded for serum samples belonging to the lung cancer patients’ group. We also explained the differences between the absorbance values of serum of healthy and lung cancer patients through a heatmap. The heatmap in Figure 6B visually differentiated between healthy and lung cancer samples, with darker colors representing high concentrations of PD-L1@EVs. The ROC curve was also established to evaluate the performance of the aptasensor with the clinical samples as shown in Figure 6C . These findings corroborate the established notion that serum from lung cancer patients harbors the greatest abundance of PD-L1@EVs than healthy controls lacking detectable PD-L1@EVs concentration (56). To validate the performance of the newly designed aptasensor, we compared it with the conventional plate ELISA for PD-L1 proposed by Chen et al. (10), using six clinical serum samples, comprising three healthy and three lung cancer samples ( Figure 6D ). Both methods showed similar results in differentiating lung cancer serum samples from healthy samples, but our method takes a shorter time than the conventional plate ELISA and does not require a separate purification step.

The proposed model (the architecture is shown in Figure 7A ) was checked for its ability to quantify the concentration of the PD-L1@EVs by analyzing the chromaticity of the tubes. We utilized a distinct confusion matrix derived from four matrices (1): true positives (TP), (2) true negatives (TN), (3) false negatives (FN), and (4) false positives (FP). TP indicates the correct identification of the actual concentration of PD-L1@EVs. TN indicates the correct identification of blank tubes without PD-L1@EVs. FN indicates incorrect prediction of PD-L1@EVs concentration as blank. FP indicates the incorrect prediction of the actual concentrations of PD-L1@EVs.

Using the values of the aforementioned four metrics, we further evaluated the following performance metrics for our deep learning model. A very straightforward metric for model evaluations is accuracy. It can be evaluated as

In Figure 7B , it can be clearly seen that our deep learning model showed excellent accuracy (99%) in predicting the concentration of PD-L1@EVs by analyzing the chromaticity of the tube. Mean absolute error (MAE) measures the performance of the model as the average magnitude of the difference between predicted values and actual values. It can be calculated as:

Here, m denotes the number of samples, yi denotes the true value, and pi denotes the predicted value. The MAE value shown in Figure 7C confirms the excellent capability of the model.

Another important graph is the precision–recall curve (PRC), which visually represents the effectiveness of a classification model at different levels of sensitivity. It assesses the relationship between two crucial aspects: precision and recall. Precision provides an assessment of the proportion of accurate positive predictions made by a model. In simpler terms, it gauges the accuracy of a model in predicting genuine positive cases. Recall, on the other hand, denotes the actual concentrations that have been accurately identified by the model ( Figure 7D ).

For a better understanding of model performance, we evaluated the F1 score. The F1 score is the harmonic mean of precision and recall. Precision, recall, and F1 score can be evaluated using the following equations, respectively:

The area under the receiver operating characteristic curve (AUROCC) is a frequently used metric for assessing a model’s ability to discriminate across classes. A model with a higher AUROC is generally considered more useful and robust ( Figure 7E ).

To further validate our model’s ability to accurately identify the concentration of PD-L1@EVs, we used 10 images for each PD-L1@EVs concentration ranging from 0 to log 10 EVs/mL to find out the performance of the deep learning model ( Figure 8 ).

As can be seen in the confusion matrix showing the relationship between the actual (true) concentration and predicted concentration of PD-L1@EVs ( Figure 8A ), 1 out of 10 tested samples at log 3 EVs/mL were misclassified as no concentration, 1 out of 10 tested samples at log 5 EVs/mL were misclassified as log 6 EVs/mL, and 2 out of 10 tested samples were misclassified as log 8 EVs/mL. The other concentration results are located diagonally in the confusion matrix, which means that they were correctly classified. Figure 8B depicts the model metrics for various concentrations of PDL1@EVs. From the exceptional metrics results, it can be inferred that the model functioned exceptionally well.

To ensure consistent evaluation of our colorimetric aptasensor readouts, we also designed a smartphone app called ExoP utilizing a CNN that can robustly predict the concentration of PD-L1@EVs by analyzing the chromaticity of the tubes from the images. The ExoP user interface can be seen in Figure 9 . This eliminates discrepancies associated with the colorimetric approach, such as lack of quantitative data without specialized equipment and inter-observer variability due to human subjectivity during visual interpretation (57). The key component behind the ExoP app is the CNN model, which has superior capabilities in gathering information from input data for learning and has been able to perform the task of distinguishing different concentrations of PD-L1@EVs from the tubes’ chromaticity on which it was not trained before. In essence, the ExoP app automates the quantitative interpretation of colorimetric aptasensor chromaticity for PD-L1@EVs concentration, making it more convenient and guaranteeing uniformity in the results. It is worth noting that the deep learning model can accurately analyze tube chromaticity, and the ExoP app can provide real-time quantitative values on the phone screen with high precision and accuracy.

Figure 10A shows that the intelligent ExoP detected a concentration of PD-L1@EVs that was quite close to the concentration detected using the aptasensor univariate analysis. Similarly, we also tested the ExoP performance using 10 clinical samples (5 healthy volunteers and 5 lung cancer patients). It can be seen in Figure 10B that there was no significant difference between the quantitative values of chromaticity revealed by the colorimetric aptasensor and the ExoP app for PD-L1@EVs.