A wearable acoustofluidic patch for rapid reversal of opioid overdose Authors: Yantao Xing, Yang Yang, Xiang Li, Hongwei Cai, Zhuhao Wu, Zheng Ao, Chunhui Tian, Vivian Niu, Jack Crystal, Gergő Szanda, Taylor J. Woodward, Heather B. Bradshaw, Han-Ting Chen, Junhua Xu, Ken Mackie, Feng Guo Published: 2026 ABSTRACT Opioids are among the most potent and reliable analgesic agents currently available, but their overdose can lead to respiratory depression and death. Although naloxone is widely used to reverse opioid overdose, it is often challenging to administer naloxone quickly enough to reverse an opioid overdose before death. Herein, we present a wearable acoustofluidic patch for rapid and automatic detection and reversal of opioid overdose. By integrating digital acoustic streaming into a 3D-printed micro-device, an acoustofluidic patch can be assembled for active and on-demand transdermal delivery of naloxone. By incorporating deep learning-guided respiratory sensing, the wearable patch can be further adapted for intelligent closed-loop treatment of opioid overdose. Moreover, our wearable patch has demonstrated successful reversal of fentanyl overdose in a mouse model, showing significantly reduced detection time, duration, and severity of overdose events, compared to the conventional method. This innovative approach may provide a promising means for personalized healthcare and digital therapy against acute disease. Introduction The misuse of opioids is a serious social crisis affecting the whole nation, negatively impacting economic welfare and public health1–3. The economic burden caused by the misuse of opioids is about $78.5 billion per year in the United States (Centers for Disease Control and Prevention). Opioid overdoses caused over 70,000 deaths in the US between June 2023 and June 2024. Prescription opioids are widely used as pain medications because they potently suppress pain perception by binding to µ-opioid receptors in the central nervous system4. However, these opioid receptors are also found in the brain regions that control respiration, and their strong activation causes severe respiratory depression, hypoxic brain injury, and even death5–8. A significant number of deaths attributed to opioids are accidental and occur during the therapeutic use of prescription opioids9 particularly when dosing is erratic10 and among workers from certain professions11. Opioid receptor antagonists such as naloxone have been developed to rapidly reverse opioid effects, including overdose12. They can restore normal respiration very quickly in most patients who have slowed or ceased breathing as a result of an opioid overdose13. Despite increasing community awareness and availability of naloxone, challenges remain to effectively rescue an individual from an opioid overdose, for several reasons: First of all, an individual is unlikely to recognize the onset of their overdose signs and symptoms14, and thus miss the opportunity to reverse it. Secondly, even during a witnessed opioid overdose, it can be extremely difficult for bystanders to administer naloxone appropriately and rapidly13,15, due to the lack of necessary medical and pharmacological knowledge and skills. Thirdly, the interval between the onset of respiratory depression and death is typically brief. Thus, the time window to seek medical help and intervene may be very short16, resulting in significant practical challenges to rescue an individual from an opioid overdose. Thus, there is an emergent and unmet need to personalize the detection and treatment of opioid overdose, especially for those who may overdose when they are alone, or who are unlikely to appropriately receive bystander-administered naloxone17–19. Several efforts have been made to develop public health solutions and engineered approaches for the reversal of opioid overdose. For individuals with opioid use disorders, virtual monitoring systems20–28, detection devices29–37, and supervised injection facilities38–41 are equipped at fixed locations to detect and treat opioid overdose. However, these public health solutions are costly and still require intensive human interventions such as manual injection of naloxone. In addition to these public health solutions, portable and automated devices and systems have been developed for personalized therapy of opioid overdose. Recently, wearable devices and sensors have been developed for real-time detection of opioid overdose by monitoring dynamic physiological parameters42–52. Moreover, prototype devices and systems have been developed for closed-loop reversal of opioid overdose through automated administration of naloxone53–60. These innovations provide promising insights for automated reversal of opioid overdose, obviating bystander intervention. However, most of these technologies rely on invasive administration techniques such as implantable devices for administering naloxone, limiting their user-friendliness for personalized therapy of opioid overdose. The rapid development of AI has revolutionized our daily lives and offers promising potential in promoting personalized therapy through advanced data handling and decision-making for wearable sensors and digital healthcare systems61–64. However, the development of AI-based devices and systems for personalized therapy of therapeutic opioid overdose is very limited. Here, we present a wearable acoustofluidic patch for the rapid detection and automated reversal of an opioid overdose (Fig. 1a). We integrate AI-based sensing and intelligent closed-loop components into a compact acoustofluidic therapeutic device that (i) detects therapeutic opioid overdose in real time via deep learning-guided respiratory sensing, (ii) triggers rapid treatment of overdose through the intelligent closed-loop system, and (iii) enables the on-demand transdermal delivery of naloxone for automated reversal of opioid overdose. Using a mouse model of opioid overdose, we demonstrated that our patch device could treat opioid-induced respiratory depression with significantly reduced detection time, duration, and severity, showing better performance than a conventional threshold-based method. This innovative technology may provide promising perspectives for personalized therapy of opioid overdose as well as other acute diseases.Fig. 1Intelligent acoustofluidic patch for rapid reversal of opioid overdose.a Illustration (left) and workflow (right) of personalized therapy for opioid overdose using our intelligent patch system supported by a cell phone-based application. The individual's respiration is monitored in real-time during a potential opioid overdose. If the deep learning algorithm runs on the cell phone and detects an overdose event, the acoustofluidic patch is triggered to reverse the patient’s respiratory depression through the digital transdermal delivery of naloxone with closed-loop regulation (indicated by blue arrows). Simultaneously, the overdose event can be reported to emergency contacts. b A schematic diagram and representative image of an intelligent acoustofluidic patch housing consisting of integrated respiratory sensors, an acoustofluidic naloxone delivery device, control circuits, and a rechargeable battery. Scale bar: 10 mm. c Working principle of the acoustofluidic transdermal naloxone delivery. Scale bar: 10 mm. Created in BioRender. Xing, Y. (2026) https://BioRender.com/ujcu8m8. Intelligent acoustofluidic patch for rapid reversal of opioid overdose. a Illustration (left) and workflow (right) of personalized therapy for opioid overdose using our intelligent patch system supported by a cell phone-based application. The individual's respiration is monitored in real-time during a potential opioid overdose. If the deep learning algorithm runs on the cell phone and detects an overdose event, the acoustofluidic patch is triggered to reverse the patient’s respiratory depression through the digital transdermal delivery of naloxone with closed-loop regulation (indicated by blue arrows). Simultaneously, the overdose event can be reported to emergency contacts. b A schematic diagram and representative image of an intelligent acoustofluidic patch housing consisting of integrated respiratory sensors, an acoustofluidic naloxone delivery device, control circuits, and a rechargeable battery. Scale bar: 10 mm. c Working principle of the acoustofluidic transdermal naloxone delivery. Scale bar: 10 mm. Created in BioRender. Xing, Y. (2026) https://BioRender.com/ujcu8m8. Results Working principle of wearable acoustofluidic patch We develop a wearable acoustofluidic patch for personal therapy of opioid overdose. Our intelligent patch system is supported by a cell phone-based application (Fig. 1a). Using a respiratory sensor, the dynamic breath of an individual can be monitored during a potential opioid overdose. If the well-trained deep learning algorithm runs on the cell phone detects a respiratory depression, the wearable patch can trigger acoustofluidic transdermal delivery of naloxone to reverse the patient’s opioid overdose with a closed-loop regulation. Simultaneously, the wearable patch can report this overdose event to emergency contacts. The compact and portable patch device (50 mm × 40 mm × 15 mm, ~30 g, Fig. 1b) consists of two respiratory sensors with sensing circuits, a transdermal naloxone delivery module with control circuits (Supplementary Fig. 1), and a rechargeable battery supporting over 48 h of continuous monitoring. The flexible respiratory sensors can be attached to the skin to track the respiration dynamics. The measured dynamic signals can be sent to the cell phone-based application for the classification of opioid overdose events. Through the closed-loop system regulation, the delivery module can be triggered for digital transdermal delivery of naloxone to reverse opioid overdose within the therapeutic window. The delivery module consists of an electromagnetic actuator, a holder, an acoustofluidic naloxone delivery device with a surface acoustic wave (SAW) generator, a naloxone capsule, and a 3D printed injector. Specifically, after receiving the trigger, the delivery module (Fig. 1c) can activate the electromagnetic actuator to penetrate the sharp tip of the injector into the skin of a human (or an animal). Meanwhile, SAW-induced streaming can pump the preloaded naloxone solution from the naloxone capsule through the microchannel of the injector for transdermal delivery with precisely and digitally controlled volume (Supplementary Fig. 2). This innovative patch prototype provides a minimally invasive, digitally-regulated, and on-demand transdermal delivery of therapeutics, enabling the rapid and automated treatment of opioid overdose with intelligent closed-loop control. We will validate and showcase its application using an animal model of opioid overdose. Acoustofluidic transdermal delivery of naloxone We fabricated and validated the acoustofluidic patch devices for transdermal delivery of naloxone using in vitro, ex vivo, and in vivo conditions. The acoustofluidic delivery patch device was fabricated using our well-developed acoustofluidics and 3D printing fabrication protocols65–67 and integrated by assembling the reusable SAW generator and the disposable injector device via epoxy resin (Supplementary Fig. 3). After applying radio frequency (RF) signals, the SAW generator excited standing SAWs, and these waves propagated along the substrate and induced acoustic streaming for pumping the liquid within the microchannel of the injector. The unique acoustic streaming induced by SAW can drive naloxone solution from the microfluidic channel/drug reservoir toward the needle outlet/transdermal tissues as a digitally regulated pump. To maximize the pumping efficiency, the microfluidic channel was designed with an incident angle of the leaky wave, the Rayleigh angle (approximately 22°)68. With the optimized design, the profile of streaming within the chamber was characterized by tracking the trajectory of microparticles (Supplementary Movie. 1). The directional jetting flows out of the injector indicate the pumping for transdermal delivery (Fig. 2a). The pumping performance of the acoustofluidic delivery patch device was also quantified with a pumping flow rate of 30 ~ 40 µL/ min (Fig. 2b). After validation in vitro, we also further tested our device for minimally invasive transdermal delivery of naloxone. The sharp needle tips were designed and fabricated using 3D printing, achieving an average tip size of 27.9 ± 2.4 µm (Supplementary Fig. 3). The hematoxylin and eosin (HE) staining results indicated the successful penetration of fresh mouse skins by the 3D-printed sharp injector (Fig. 2c). Moreover, our results also showed rapid skin recovery in vivo after 6 h of the injector treatment, indicating that transdermal delivery using the device is minimally invasive (Supplementary Fig. 4). Furthermore, different from syringe injection, the acoustofluidic patch mediated delivery of Rhodamine B ex vivo demonstrates uniform lateral dispersion within the dermis, without formation of a confined subcutaneous pocket (Supplementary Fig. 5). Then, the device was further validated for controllable transdermal delivery of naloxone.Fig. 2Acoustofluidic transdermal delivery of naloxone with excellent dosage control.a Schematics (left) showing an innovative delivery mechanism of our acoustofluidic patch based on acoustic streaming induced by surface acoustic wave (SAW). Stacked fluorescence images (right) describing the streaming lines without (SAW OFF) or with (SAW ON) streaming excitation (Vpp = 17 V, frequency = 37.4 MHz, and duty cycle = 50%). Jetting flows are labeled by red arrows, indicating the liquid pumping for transdermal delivery. Scale bar: 1 mm. b Quantification of pumping volumes over time under SAW ON and OFF conditions (mean ± sd, n = 5 independent measurements per group). c Hematoxylin and eosin (H&E) staining of the mouse skin with or without acoustofluidic patch penetration. Scale bar: 500 µm. Representative images are shown from three independent skin samples with similar results. d Quantification of transdermally delivered naloxone over time ex vivo (mean ± sd, n = 3 independent samples from different mice per time point). Statistical significance was determined using a two-sided unpaired Student’s t-test for the indicated comparisons; **P < 0.01. Sham controls (needle engaged, SAW off) were measured using the same protocol, and all values were below the detection limit. e Comparison of transdermally delivered naloxone using acoustofluidic patch administration and subcutaneous injection with a target dose of 1.5 mg (black dotted line), (mean ± sd, n = 3 independent samples from different mice per group). f Bio-distribution (e.g., serum, brainstem, and liver) of acoustofluidic patch-mediated transdermally delivered naloxone in vivo (mean ± s.d., n = 4 independent mice per time point). Acoustofluidic transdermal delivery of naloxone with excellent dosage control. a Schematics (left) showing an innovative delivery mechanism of our acoustofluidic patch based on acoustic streaming induced by surface acoustic wave (SAW). Stacked fluorescence images (right) describing the streaming lines without (SAW OFF) or with (SAW ON) streaming excitation (Vpp = 17 V, frequency = 37.4 MHz, and duty cycle = 50%). Jetting flows are labeled by red arrows, indicating the liquid pumping for transdermal delivery. Scale bar: 1 mm. b Quantification of pumping volumes over time under SAW ON and OFF conditions (mean ± sd, n = 5 independent measurements per group). c Hematoxylin and eosin (H&E) staining of the mouse skin with or without acoustofluidic patch penetration. Scale bar: 500 µm. Representative images are shown from three independent skin samples with similar results. d Quantification of transdermally delivered naloxone over time ex vivo (mean ± sd, n = 3 independent samples from different mice per time point). Statistical significance was determined using a two-sided unpaired Student’s t-test for the indicated comparisons; **P < 0.01. Sham controls (needle engaged, SAW off) were measured using the same protocol, and all values were below the detection limit. e Comparison of transdermally delivered naloxone using acoustofluidic patch administration and subcutaneous injection with a target dose of 1.5 mg (black dotted line), (mean ± sd, n = 3 independent samples from different mice per group). f Bio-distribution (e.g., serum, brainstem, and liver) of acoustofluidic patch-mediated transdermally delivered naloxone in vivo (mean ± s.d., n = 4 independent mice per time point). We characterized the transdermally delivered naloxone by the device ex vivo and in vivo using high-performance liquid chromatography-mass spectrometry (HPLC-MS). By tuning the treatment time, the transdermally delivered naloxone within the mouse tissues was measured (Fig. 2d), and a delivery rate was calculated as 0.55 ± 0.05 mg/min, whereas all sham (needle-engaged, SAW-off) samples yielded naloxone levels below the detection limit. Notably, the delivered dose increased gradually with treatment time, consistent with a diffusion-dominated transdermal transport process rather than pressure-driven bolus injection. To achieve the on-demand dosage control, a target dosage of transdermally delivered naloxone (e.g.,1.5 mg, the dosage to rescue a mouse) was delivered via our device. The results showed that the acoustofluidic delivery patch device achieved comparable accuracy and robustness to the syringe-based method, with doses more closely aligned to the 1.5 mg target. (Fig. 2e). To further investigate the pharmacokinetics, the biodistribution of the transdermally delivered naloxone was characterized in vivo over time. The results showed the concentration variations of transdermally delivered naloxone in different organs and samples (e.g., serum, brainstem, and liver) post-treatment (Fig. 2f). The transdermally delivered naloxone was detectable in all tissues within 5 min post-treatment. In the sham patch condition, naloxone levels remained at or near the detection limit across all tissues, indicating minimal delivery in the absence of acoustic actuation. Although absolute concentrations differed due to distinct pharmacokinetic mechanisms, the temporal serum concentration profiles observed with transdermal delivery were consistent with those obtained from subcutaneous injection under matched experimental conditions (Supplementary Fig. 6a), and aligned with previously reported early-time profiles of parenteral naloxone administration69,70. Due to the brainstem’s high lipid content, peak naloxone levels were approximately five times higher in the brainstem compared to the liver, indicating the potential in treating opioid overdose through the acoustofluidic transdermal delivery of naloxone. Open-loop reversal of opioid overdose To further test the acoustofluidic transdermal delivery of naloxone for rescuing opioid overdose, we developed an acoustofluidic patch device-based open-loop system and validated the effectiveness of the patch device and system using a mouse model of opioid overdose (Fig. 3a). The dynamic respiratory pattern of a healthy mouse was tracked in real-time via the flexible respiratory sensors or the whole-body plethysmography (WBP). Through the induction of opioid overdose (at Tinduction), the respiration of the mouse administered overdosed fentanyl at a dose of 0.4 mg/kg was continuously monitored. Once the respiratory depression is detected (at Toverdosed), the open-loop system triggers the acoustofluidic transdermal delivery of naloxone with a desired dosage of 1.5 mg. The respiration of the overdosed mouse was continuously monitored until they fully recovered from their overdosed condition. To optimize the identification of respiratory depression, representative lung flow changes over time were recorded during the induction and reversal of opioid overdose (Fig. 3b). Additional ventilatory parameters, including respiratory rate, tidal volume, and minute volume, were calculated from the lung flow changes. From the switching from healthy to overdosed conditions, these three ventilatory parameters were all significantly changed, while the minute volume presented the most significant changes (Fig. 3c). Because the minute volume combines the changes of both respiratory frequency and tidal volume, representing the total physiological airflow into and out of the lungs (Supplementary Fig. 7)71. In mouse models of opioid-induced respiratory depression, minute ventilation is a widely used surrogate endpoint for respiratory depression and its reversal in awake, freely moving animals72. Thus, the minute volume threshold-based detection was selected as the criterion for the identification of respiratory depression. With these optimized conditions, the acoustofluidic patch device-based open-loop system was validated using the mouse model of opioid overdose. The representative minute volume dynamics of the mouse post-induction of opioid overdose first showed the detection of overdose (red circles) and the reversal of overdose (black circles) through the acoustofluidic transdermal delivery of naloxone, while the minute volume of the mouse post-induction but without any treatment remained below 50% of baseline (Fig. 3d). Thus, the results on the open-loop reversal of opioid overdose indicated the feasibility to detect the onset and recovery of opioid overdose and to treat the opioid overdose through acoustofluidic transdermal delivery of naloxone.Fig. 3Acoustofluidic naloxone delivery for rescuing mice from opioid overdose.a Schematic procedure showing the induction, detection/treatment, and recovery of opioid (fentanyl) overdose in mice. b Representative box flow graphs describing breathing patterns of a mouse in healthy, overdosed, and recovered conditions. c Comparison of three respiratory parameters between healthy and overdosed conditions, with the most pronounced decrease observed in minute volume (mean ± sd, n = 5 mice per condition). Statistical significance was determined using an unpaired Student’s t-test; *P < 0.05, **P < 0.01 and ***P < 0.001. d Representative minute volume tracings with or without acoustofluidic patch treatment, where the red circles represent the duration of opioid overdose. Traces were smoothed using a three-point moving average. Acoustofluidic naloxone delivery for rescuing mice from opioid overdose. a Schematic procedure showing the induction, detection/treatment, and recovery of opioid (fentanyl) overdose in mice. b Representative box flow graphs describing breathing patterns of a mouse in healthy, overdosed, and recovered conditions. c Comparison of three respiratory parameters between healthy and overdosed conditions, with the most pronounced decrease observed in minute volume (mean ± sd, n = 5 mice per condition). Statistical significance was determined using an unpaired Student’s t-test; *P < 0.05, **P < 0.01 and ***P < 0.001. d Representative minute volume tracings with or without acoustofluidic patch treatment, where the red circles represent the duration of opioid overdose. Traces were smoothed using a three-point moving average. Incorporation of deep learning-guided respiratory sensing for the detection of opioid overdose To achieve automated detection and further intelligent closed-loop control, we incorporated a deep learning model to classify respiratory depression based on dynamic respiratory features. This approach not only extracts time-domain characteristics but also captures temporal dependencies. Data for the model training and validation were collected from ~780 respiratory datasets from a total of 78 mice (Fig. 4a). All signals were calibrated and normalized to ensure the stability of the extracted features. The minute ventilation was calculated from the respiratory data and normalized as the input for the model. A sliding window method was applied to the data, with window sizes ranging from 1 to 10 min. To perform the binary classification task of distinguishing between overdose and healthy baseline states, a convolutional neural network (CNN) with three one-dimensional convolutional layers was implemented. During training, 50% of the training data was reserved for model validation. We first compared the model accuracy for different sliding window lengths and determined that a 5-min window provided the optimal training configuration (Fig. 4b). The results from the testing dataset are summarized in the confusion matrix (Fig. 4c), which demonstrated an accuracy of 0.95 and an F1 score (assessing a model’s overall performance in classification tasks) of 0.948. Furthermore, the performance of the CNN model was compared with other machine learning models: a recurrent neural network (RNN) and an artificial neural network (ANN). For comparison, a threshold-based method (‘threshold’, as minute volume below 50% of baseline) was also included. The receiver operating characteristic (ROC) curves (Fig. 4d) indicated that all three models achieved an area under the curve (AUC) greater than 0.90, significantly outperforming the conventional threshold method and thereby confirming the effectiveness of the proposed framework. Notably, the CNN model achieved the best performance with an AUC of 0.986, enabling real-time decision making for closed-loop treatment.Fig. 4Deep learning-guided respiratory sensing for detection of opioid overdose in mice.a Schematics describing deep learning algorithms (Convolutional Neural Network, or CNN) for opioid overdose classification using a training dataset of respiratory traces from mice. b Dependence of model accuracy upon the sliding window. Data are presented as mean ± s.d., n = 5 cross-validation folds. c A confusion matrix indicating results from the test group using data collected with/without opioid overdose induction over up to 40 min, achieving an overall accuracy of 95.0%. d Comparison of the receiver operating characteristic (ROC) curves with models including CNN, recurrent neural network (RNN), artificial neural network (ANN), threshold-based algorithm, and random chance, indicating the best performance of the CNN model. e The representative respiratory curve of a dynamic opioid overdose over time indicates that deep learning-based detection (light red circles) can detect an opioid overdose earlier than the threshold-based method (dark red circles). f Comparison of detection times between deep learning and threshold-based detections. Each color indicates an individual mouse (n = 5). Statistical significance was determined using a two-sided paired Student’s t-test; **P < 0.01. Deep learning-guided respiratory sensing for detection of opioid overdose in mice. a Schematics describing deep learning algorithms (Convolutional Neural Network, or CNN) for opioid overdose classification using a training dataset of respiratory traces from mice. b Dependence of model accuracy upon the sliding window. Data are presented as mean ± s.d., n = 5 cross-validation folds. c A confusion matrix indicating results from the test group using data collected with/without opioid overdose induction over up to 40 min, achieving an overall accuracy of 95.0%. d Comparison of the receiver operating characteristic (ROC) curves with models including CNN, recurrent neural network (RNN), artificial neural network (ANN), threshold-based algorithm, and random chance, indicating the best performance of the CNN model. e The representative respiratory curve of a dynamic opioid overdose over time indicates that deep learning-based detection (light red circles) can detect an opioid overdose earlier than the threshold-based method (dark red circles). f Comparison of detection times between deep learning and threshold-based detections. Each color indicates an individual mouse (n = 5). Statistical significance was determined using a two-sided paired Student’s t-test; **P < 0.01. To apply and validate the deep learning model for detecting opioid overdose, we conducted deep learning-guided respiratory sensing experiments and compared their performance with the threshold-based method (Fig. 4e). Immediately after the induction of opioid overdose, the representative dynamic respiratory curve didn’t show any noticeable changes. Approximately 12 min post-induction, the deep learning-guided respiratory sensing method was able to detect the impending overdose event, while after ~8 min, the same event was not detected by the conventional threshold-based respiratory sensing method (current method for detecting an overdose in a clinical setting). Moreover, the quantification results further showed that the deep learning-based method can detect overdose events earlier than the threshold-based method by about 11 min (Fig.4f and Supplementary Fig. 8). Deep learning-guided respiratory sensing enables the early detection and timely administration of naloxone and may significantly improve treatment efficacy and efficiency. Intelligent closed-loop reversal of opioid overdose in mice We further developed an intelligent closed-loop system for the reversal of opioid overdose (Fig. 5a). This intelligent patch system includes respiratory sensing, acoustofluidic transdermal delivery of naloxone, and a cell phone-based deep learning application (Supplementary Movie. 2). One hour before the experiment, the mice were fitted with wearable acoustofluidic patch devices to allow for adaptation and minimize potential interference with respiratory signals due to agitation. The dynamic respiratory data were continuously measured and recorded at a refresh rate of 30 s. After achieving the stabilized measurement of respiratory signals, the mice were administered fentanyl for the induction of opioid overdose. The intelligent patch system was controlled by a smartphone with a well-trained deep learning algorithm for real-time detection of opioid overdose and initiation of acoustofluidic transdermal naloxone delivery. Representative dynamic minute volume curves showed the respiratory depression and recovery with the intelligent closed-loop treatment and the threshold-based closed-loop treatment (Fig. 5b). Very importantly, the intelligent closed-loop treatment approach demonstrated the earlier detection and treatment, as well as the shorter duration and severity than the threshold-based closed-loop treatment. It is worth noting that the intelligent and threshold approaches used the same closed-loop system and the same sensors and acoustofluidic patch devices, while the difference is that one is triggered via a deep learning method, but the other is triggered via a conventional threshold value. Moreover, we further quantitatively compare both methods using 6 mice per method, respectively (Supplementary Figs. 9 and 10). Our quantification results confirmed that the intelligent system significantly reduced the duration of overdose by an average time of ~11 min (Fig. 5c) and the severity (Fig. 5d) of the overdose events, compared with the conventional threshold-based strategy73–76. In addition, under matched naloxone administration timing, the patch achieved respiratory recovery comparable to manual i.v. naloxone administration (Supplementary Fig. 11). These results indicate the faster and safer treatment of the intelligent patch system than the threshold-based strategy.Fig. 5Intelligent patch for automated and rapid reversal of opioid overdose in mice.a Flowchart illustrating the rapid detection and intelligent closed-loop treatment of opioid overdose in mice. b Representative respiratory curves during the reversal of opioid overdose in mice using intelligent (deep-learning algorithms) and threshold (50% decrease of minute volume value)-based closed-loop systems, where both systems use acoustofluidic transdermal delivery of naloxone. Arrows indicate the time points of overdose detection by deep learning and threshold-based methods, red circles label the time points of ‘overdose duration’, and the gray area between the dashed baseline and the curve of red circles marks ‘severity area’. Comparison of the overdose duration (c) and severity area (d) of the intelligent and threshold-based systems (mean ± sd, n = 6 mice per group), indicating the faster and safer treatment of the intelligent system than the threshold-based system. Statistical significance was determined using a two-sided unpaired Student’s t-test; *P < 0.05, **P < 0.01 and ***P < 0.001. Intelligent patch for automated and rapid reversal of opioid overdose in mice. a Flowchart illustrating the rapid detection and intelligent closed-loop treatment of opioid overdose in mice. b Representative respiratory curves during the reversal of opioid overdose in mice using intelligent (deep-learning algorithms) and threshold (50% decrease of minute volume value)-based closed-loop systems, where both systems use acoustofluidic transdermal delivery of naloxone. Arrows indicate the time points of overdose detection by deep learning and threshold-based methods, red circles label the time points of ‘overdose duration’, and the gray area between the dashed baseline and the curve of red circles marks ‘severity area’. Comparison of the overdose duration (c) and severity area (d) of the intelligent and threshold-based systems (mean ± sd, n = 6 mice per group), indicating the faster and safer treatment of the intelligent system than the threshold-based system. Statistical significance was determined using a two-sided unpaired Student’s t-test; *P < 0.05, **P < 0.01 and ***P < 0.001. Discussion We have developed a wearable acoustofluidic patch that enables rapid and on-demand transdermal delivery of naloxone, providing personalized and responsive treatment for opioid overdose. This device integrates digital acoustic streaming and a 3D-printed micro-device to achieve precise and active drug delivery. Operating in a delivery regime distinct from diffusion-limited transdermal systems, the patch enables active liquid transport rather than passive molecular permeation. By leveraging deep learning-guided respiratory sensing, the patch was adapted for closed-loop opioid overdose treatment, ensuring real-time detection and intervention. As a proof-of-concept, our device demonstrated faster respiratory recovery from opioid overdose, with reduced detection time and respiratory suppression severity in a preclinical model compared with conventional threshold-based naloxone therapy. This acoustofluidic patch addresses key limitations of existing transdermal and wearable drug-delivery systems. Conventional transdermal platforms rely primarily on passive diffusion across the skin barrier, leading to delivery kinetics that are difficult to modulate in real time and sensitive to inter-individual variability in skin properties. In contrast, this acoustofluidic patch device enables active, programmable liquid transport across the skin through SAW-induced streaming, allowing dosing to be dynamically regulated rather than solely diffusion-driven. Clinically, the significance of this capability lies not in replacing established naloxone formulations but in addressing unmet needs related to timing, autonomy, and individual variability in opioid overdose intervention. By integrating physiological sensing with automated, on-demand transdermal delivery, the platform introduces a responsive and individualized intervention paradigm that may complement existing rescue strategies, particularly for high-risk populations or settings where immediate bystander intervention is unreliable. This innovative platform holds promise for personalized therapy of opioid overdose, pain management, and other acute conditions. Future iterations of the intelligent acoustofluidic patch could benefit from improvements in wearability, personalized adaptation, and clinical translation. Optimizing material structures and ergonomic design may enhance skin compatibility, flexibility, and long-term comfort, making the device more suitable for practical use. Additionally, incorporating continuous learning mechanisms could enable the system to adapt to individual physiological patterns over time, optimizing drug release based on personalized biometric responses. Such an approach may further improve treatment precision and safety. Furthermore, our in vivo validation in a mouse overdose model serves as a proof-of-concept and was not designed as a head-to-head efficacy comparison with existing clinical formulations/routes of administration of naloxone. While intranasal and intramuscular naloxone represent the current standard of care for bystander rescue, rigorous route-to-route comparisons would require formulation-specific optimization, matched pharmacokinetic characterization, and translationally relevant animal models. In addition, interspecies differences in respiratory physiology, skin barrier properties, and naloxone clearance limit direct extrapolation from mice to humans. To bridge this gap, large-animal studies with systematic evaluation of pharmacokinetics, safety, and reproducibility will be essential. In parallel, transfer learning and cross-species domain adaptation will enable refinement on human respiration data, while multi-class modeling and artifact-robust, multimodal sensing will address etiologic differentiation and degraded signals. Methods Fabrication of acoustofluidic patch devices The acoustofluidic patch devices were developed through three steps, including 3D printing of patch devices (with a hollow injector and streaming channel at the microscale), fabrication of a surface acoustic wave (SAW) generator, and assembly of these two components into an integrated device. First, the patch devices with a hollow injector and streaming channel were designed using AutoDesk Fusion 360 (Autodesk Inc, San Rafael, CA, USA), fabricated using a stereolithography (SLA) 3D printer (Sonic Mini 8 K, Hsinchu, Taiwan), and characterized using a stereomicroscope (Leica M205FA, Supplementary Fig. 1). Moreover, the SAW generator was made by fabricating the Cr/Au electrode patterns onto a 128° Y-cut lithium niobate (LiNbO₃) wafer using photolithography and metal deposition procedures77–80. Furthermore, the acoustofluidic therapeutic patch device was assembled by directly bonding the 3D-printed device onto the SAW generator using double-sided tape. Operation and characterization of acoustofluidic patch devices The assembled acoustofluidic patch devices were driven by radio frequency (RF) signals (sine wave; frequency = 37.4 MHz; duty cycle = 50%) to achieve on-demand transdermal delivery using RF excitation with the parameters described above81–85. To characterize the acoustic streaming in the patch device, the trajectories of fluorescent 2 µm polystyrene beads were recorded using a microscope (IX-81, Olympus, Japan) and analyzed using ImageJ 1.53t. For visualization purposes, a modified injector device was fabricated in which one lateral side of the microchannel was left open and sealed with a thin glass coverslip using epoxy, providing optical access while preserving the internal channel geometry. A suspension of 2 µm fluorescent polystyrene beads (100 × dilution; Bangs Laboratories) was introduced into the microchannel. Particle motion was recorded under RF-off and RF-on conditions using a fluorescence microscope (IX-81, Olympus). Steady-state acoustic streaming patterns were obtained by overlaying sequential image frames to extract particle trajectories. To quantify the pumping performance, microtubing was connected to the inlet of the acoustofluidic patch device, allowing liquid to enter the device from an external reservoir. After removing all air bubbles from the microneedle and tubing, the tubing inlet was positioned at the same vertical height as the needle tip to eliminate gravitational or siphon-driven flow. Pumped volume was determined by tracking the displacement of the air–liquid interface within the tubing and multiplying the traveled distance by the internal cross-sectional area. The flow rate was calculated by dividing the pumped volume by the actuation duration. For the in vitro experiments, the device orifice was penetrated into an agar phantom to mimic the tissue resistance. Animals All experimental procedures were carried out in compliance with the National Institutes of Health Guidelines for the Care and Use of Laboratory Animals and approved by the Institutional Animal Care and Use Committees (IACUC) at Indiana University Bloomington (protocol# 21-033). All efforts were made to minimize the suffering and number of animals used. Adult male and female wild-type outbred CD-1 IGS mice (Mus musculus; Crl:CD1(ICR), Charles River Laboratories, stock no. 022), weighing 25–35 g and 12–18 weeks of age, were obtained from a breeding colony maintained at Indiana University Bloomington. Mice were housed in HEPA-filtered air at 20–24 °C and 30–70% relative humidity under a 12-h light/dark cycle, with lights on from 07:00 to 19:00. Food and water were available ad libitum. Ex vivo and in vivo validation of transdermal naloxone release Our acoustofluidic patch devices loaded with naloxone were applied to freshly harvested skin tissues for varying durations of treatment. The penetration of the stratum corneum by the injector was characterized by cryosection and H&E staining86. Transdermal delivery of naloxone was quantified by high-performance liquid chromatography-mass spectrometry (HPLC-MS). The biodistribution of transdermally delivered naloxone was also characterized by HPLC-MS/MS. After 3 min of transdermal naloxone delivery via our acoustofluidic patch devices, the mice were sacrificed at different time points (5, 10, 15, 30 min), and blood, brainstem, and liver samples were collected and analyzed. For ex vivo visualization of acoustofluidic patch-mediated transdermal delivery, Rhodamine B (RhB, 10 mg/mL) was prepared in 1× PBS and used as a fluorescent tracer. Freshly harvested mouse skin tissues were mounted with the acoustofluidic patch and treated with different acoustic actuation durations. After treatment, the skin surface was wiped with acetone to remove residual surface dye and then imaged using a fluorescence microscope (IX83, Olympus). Bioanalytical quantification of naloxone by HPLC-MS/MS Naloxone concentrations in serum and tissue samples were quantified using an analytical workflow of LC–MS/MS87. Brainstem and liver tissues were rapidly dissected, flash frozen, and stored at −80 °C until analysis. On the day of extraction, samples were weighed and homogenized in HPLC-grade methanol, with solvent volumes scaled to tissue mass. An internal standard was added prior to extraction to enable assessment of extraction efficiency. Following homogenization, samples were centrifuged, and supernatants were diluted with water prior to partial purification using C18 solid-phase extraction columns. Sequential methanolic elution was performed, and defined elution fractions were collected. Aliquots of the final extracts were injected for LC–MS/MS analysis. Quantification was performed against external calibration standards prepared from authentic naloxone. Measured signal intensities were converted to molar amounts using calibration curves and back-calculated to concentrations by accounting for dilution factors, elution volume, injection volume, and tissue mass. Concentrations are reported as µg/mL for serum samples and µg/g for tissue samples. Extraction recovery was calculated on a per-sample basis using the internal standard, with observed recovery values ranging from approximately 35% to 80% across tissue samples. Recovery-adjusted concentrations were used for all quantitative analyses, and no systematic matrix-dependent signal suppression or enhancement was observed following recovery correction. The lower limit of quantification (LLOQ) of the assay was below the lowest measured concentrations, and the upper limit of quantification (ULOQ) exceeded the highest observed concentrations, ensuring that all reported values fell within the validated linear range of the assay. Validation of patch devices using a mouse model of opioid overdose As a mouse model of opioid overdose, the mice were administered a single dose of fentanyl citrate (0.4 mg/kg, Hikma, London, UK). The fentanyl dose was selected based on prior studies demonstrating its efficacy in inducing severe respiratory depression and loss of righting reflex in mouse models88. Before administration, mice were acclimated to the experimental environment for at least 30 min to minimize stress-induced physiological variability. The patch was affixed to the dorsal skin under brief isoflurane anesthesia (1–2% in oxygen, <2 min) to ensure proper attachment and uniform drug diffusion. Isoflurane was then discontinued, and animals were allowed to fully recover (return of righting reflex and spontaneous locomotion) before the experiment proceeded. Mice were maintained in room air with the patch in place for ~1 h to allow complete washout of any residual anesthetic and to ensure stable baseline respiration. Throughout the fentanyl infusion and post-administration period, real-time respiratory sensing was performed using whole-body plethysmography (WBP) or our soft sensors. Following fentanyl exposure, a fixed dose of naloxone (1.5 mg) was given through the patch over 3 min to reverse the opioid overdose. This dose is consistent with prior reports that 1–2 mg naloxone effectively rescues fentanyl-induced respiratory depression in mice89. The whole opioid overdose process lasted for about 120 min, and all mice were sacrificed after the experimental procedures. Integration, training, and validation of intelligent patch devices The intelligent patches were integrated by incorporating deep learning-guided respiratory sensing for detecting opioid overdose and closed-loop triggering of acoustofluidic patch naloxone delivery for managing opioid overdose. The deep learning model was trained and validated for detecting opioid overdose using the respiratory data of mice. Animal respiratory data were collected from a total of 78 individual mice, capturing respiratory patterns under both baseline (healthy) and post-overdose induction conditions. The raw data were segmented into fixed-length windows of 5 min, ensuring uniform input dimensions for model training. For training purposes, data were manually labeled as “healthy” or “post-overdose induction”. A convolutional neural network (CNN) was employed to classify respiratory patterns and predict overdose events. The CNN architecture comprised three alternating convolutional layers, each followed by a batch normalization and rectified linear unit (BN-ReLU) activation function to improve feature extraction and model stability. The softmax activation function was utilized in the output layer to generate the final classification probabilities, assigning each 5-min data window a probability of being in an overdose state. To enable real-time overdose monitoring and automated therapeutic intervention, a Python-based (Python 3.10) control script was developed for deep-learning model inference. This script continuously monitored the classification output with a 30-s refresh rate, ensuring near real-time detection of respiratory depression events. When the model classified an input window as an overdose (classification probability = 1), the script automatically triggered the activation of the acoustofluidic patch, initiating naloxone transdermal delivery for reversing overdose conditions. Ethics All animal experiments were conducted in accordance with protocols approved by the Institutional Animal Care and Use Committee of Indiana University Bloomington (IACUC protocol no. 21-033). Statistical analysis The statistics comparing different testing groups were conducted using the t-test. Statistical significance was denoted as follows: *P < 0.05, **P < 0.01, ***P < 0.001. Reporting summary Further information on research design is available in the Nature Portfolio Reporting Summary linked to this article. Supplementary information Supplementary Information Supplementary Movie 1: Acoustic streaming for digitalized drug delivery. Supplementary Movie 2: Smartphone App for potential personal therapy of opioid overdose. Reporting Summary Transparent Peer Review file Supplementary Information Supplementary Movie 1: Acoustic streaming for digitalized drug delivery. Supplementary Movie 2: Smartphone App for potential personal therapy of opioid overdose. Reporting Summary Transparent Peer Review file Source data Source Data Source Data