## Introduction

The human body responds to injury with bleeding, followed by clot formation and eventually lysis1. This carefully maintained homeostasis minimizes the risks of hemorrhage and inappropriate clotting like ischemic stroke, myocardial infarction or pulmonary embolus2. However, for millions of people, medical conditions such as atrial fibrillation, mechanical heart valves and genetic mutations increase the risk of morbidity and mortality from blood clotting2. These individuals require lifelong administration of anticoagulation drugs such as warfarin, an effective medication but also one of the most common causes of hospitalization due to adverse drug events3. Hence, medication effects must be closely monitored via frequent prothrombin time (PT) or international normalized ratio (INR) tests to assess coagulation properties due to the drug’s narrow therapeutic index and interactions with food and other medications4. While newer anticoagulants that do not rely on regular PT/INR testing are increasing in popularity, studies show that warfarin remains the most commonly prescribed outpatient blood thinner5.

PT/INR testing monitors extrinsic and common pathways of the coagulation cascade. These tests are usually performed in a laboratory on expensive equipment after separating plasma from whole blood. Home PT/INR monitors directly utilize blood and have been shown to lengthen the amount of time spent in the therapeutic range, decrease the risk of thromboembolism in young patients and improve patient satisfaction and quality of life6,7. Time spent in the therapeutic window benefits patients with non-valvular atrial fibrillation on warfarin, since the risk of bleeding is five times higher with overly aggressive anticoagulation and the risk of ischemic events is three times higher with insufficient anticoagulation compared to levels in the therapeutic range8.

Despite the existence of several home PT/INR testing modules, access to affordable and accurate PT/INR testing remains a challenge. Patients in the United States are in the therapeutic range only about 64% of the time6,9. Patients in developing countries like Botswana, Uganda, and India are in this range, only 40% of the time due to less frequent testing9,10,11. Despite the potential for improvements from home PT/INR testing9, these devices cost hundreds of dollars, limiting their utility in resource-constrained environments12,13.

Here, we describe a proof-of-concept system that uses the vibration motor and camera on existing smartphones to perform PT/INR testing. Smartphones are increasingly becoming ubiquitous in resource-constrained environments and developing countries both in rural and urban settings14,15,16. Vibration motors and cameras have been an integral part of smartphones for more than a decade. Repurposing these smartphone sensors for PT/INR testing could enable a more affordable blood clot testing tool.

Our system visually tracks the micro-mechanical movements of a small copper particle in a cup with either a single drop of whole blood or plasma and the addition of activators. No additional electronic components are required beyond a lightweight plastic attachment that couples the phone’s vibrations to the cup. When the mixture is in a fluid state, the copper particle moves freely with the phone’s vibration. As the blood clots, the viscous mixture causes the particle to slow to a stationary state. Using computationally efficient video analytic algorithms that run on the smartphone in under 37 ms, we analyze the particle’s motion in under one minute to determine the PT/INR values. Our system can run on older smartphones such as second-hand iPhone 5s phones that were released in 2013 and cost $35. Making coagulation testing accessible in this manner may help improve time within the therapeutic range for anti-coagulation users, particularly in rural locations. ## Results ### Concept and prototype Our design builds on conventional central laboratory and point-of-care coagulation tests. Since manual visual detection of clot time is subjective, requires training and can vary between operators and institutions17, existing tests use optical and magnetic Hall-effect sensors to automatically track changes in plasma viscosity and turbidity18. Mechanical approaches use specialized hardware to analyze the movements of steel balls19, iron fillings20, and magnets21 in the plasma sample. Other automated approaches include electrochemical sensors that measure capacitance and resistance22,23,24,25, quartz crystal resonators26, micro-resonators18, and centrifugal-type microfluidic platforms27. Though laboratory clotting assays enable high-throughput testing, they may not be able to provide rapid turnaround for anticoagulant therapy in emergency room and intensive care units, which often require results within 30 minutes and around-the-clock availability28. Commercial point-of-care systems eliminate these delays but can be expensive for field use, in resource-constrained settings and for in-home patients, who cite cost of PT/INR home devices as the primary barrier to self-testing13. Unlike these conventional coagulation tests, our design is low-cost, with a total material cost of around$0.03, and requires only a lightweight (17 g), compact (70 × 27.5 × 60.9 mm), 3D-printed plastic smartphone attachment, disposable plastic cup, and tiny copper particle. Further, our solution leverages on-board vibration motors and cameras that are ubiquitous on modern smartphones and are accessible in resource-constrained environments. Finally, ours is an automated solution that does not require manual observation or interpretation of clotting data.

Our system leverages the smartphone’s onboard vibration motor to vibrate a small silicone cup (Fig. 1). The smartphone is coupled to the cup via a custom 3D-printed plastic attachment. The ‘L’ shaped structure of the plastic attachment is constructed using thin material to allow the smartphone’s vibrations to propagate to the cup holder while reducing dampening. In addition, the cup holder makes physical contact with both the bottom and sides of the cup to ensure maximal physical energy transfer from the attachment to the cup. The plastic attachment is designed to position the cup under the smartphone’s camera and its width is determined by the smartphone’s dimensions.

The cup holds 10–20 μl of plasma or whole blood and a small copper particle, provided as part of our design (Fig. 1). A thromboplastin activator is added to this mixture to activate the extrinsic pathways of the coagulation cascade. The copper particle is a 1 mm long AWG22 copper wire coated with dark blue ink for increased visibility. It is lightweight and moves freely in response to the phone’s vibration. In addition, it is non-porous, which prevents it from soaking up any reagent. We chose copper material because it is also low-cost, widely available, and easily cut to short lengths. The form factor of our proof-of-concept design could be further optimized for patient use by integrating the copper particle into single-use disposable cups that are coated with a dried form of thromboplastin29,30.

When the blood or plasma sample is not coagulated, the smartphone’s vibrations cause the copper particle to move and rotate within the sample. As the sample coagulates, its increasing viscosity constricts the particle’s movements, slowing its projected 2D motion as seen by the camera (see Supplementary Movie 1). The particle’s movement is recorded optically by the smartphone camera and analyzed to calculate PT/INR. In particular, our algorithm identifies when the activator is added and computes the coagulation time automatically. We isolate the 2D movement of the particle from the background and perform a correlation analysis of video frames to generate motion curves (Fig. 2). The start of a measurement is detected when the activator is dispensed to the cup. A steep drop in the particle’s motion curve indicates the particle’s stationary state and marks the mixture’s clot time. These motion curves are then analyzed to algorithmically identify the start and end time points of the measurement and compute the PT/INR values (see the “Methods” section). Internal checks are performed before, during, and after a measurement to ensure the test has been performed properly (Supplementary Table 2). If all internal checks have passed, the PT/INR result is displayed.

As our system only processes video frames captured every 100 ms to calculate PT/INR, any smartphone camera that can capture video at 10 fps or greater is able to record the clotting process. All modern smartphones capable of capturing video records at a minimum of 24 or 30 fps, which is the standard frame rate for video recordings. We note also that when the video resolution was downsampled to 960 × 540 there was no change in computed PT/INR values for plasma and blood samples, which shows that lower resolution smartphone cameras can capture the clotting process. The most common resolution for video recording on modern smartphones is 1920 × 1080.

### Clinical testing on plasma samples

We evaluated our system using 140 anonymized plasma samples from the University of Washington Medical Center (UWMC). A subset of these samples included plasma from patients who were undergoing treatment at the UWMC Anticoagulation clinic and were thus likely to have high PT/INR values. Samples were marked with PT/INR values obtained from a clinical-grade coagulation analyzer (Diagnostica Stago STA R Max). They were collected and tested on our smartphone setup within 12 h of being drawn from the patients as prior studies have shown no significant change in PT values for samples up to 24 h from blood draw31,32. The plasma samples were stored at room temperature prior to measurement on our system. The PT/INR values of the plasma samples ranged from 11.4–49.8 s and 0.9–5.3, respectively, with a mean value of 20.1 s and 1.8 and a median value of 17 s and 1.4 (Supplementary Fig. 2a).

Since the smartphone measurements were performed hours after the plasma samples were collected, the samples (and the thromboplastin activator) were heated in a water bath at 37 °C (approximating the body temperature) for three minutes. The measurement was then conducted at room temperature and humidity. As the test takes less than a minute, ambient temperature and humidity within normal limits do not significantly affect the performance of the system (see Supplementary Table 1 and Supplementary Materials for details). Twenty microliters of the plasma was added into the cup with the copper particle and placed into the smartphone attachment. The smartphone’s vibration motor was turned on to vibrate continuously, and the camera began recording. Forty microliters of the activator was then added into the cup. Samples were tested on a Samsung Galaxy S9 phone and each test was performed twice to evaluate test–retest performance.

The PT/INR values computed by the smartphone system were compared against the laboratory PT/INR values. The inter-class correlation coefficient for PT/INR was R = 0.963 and R = 0.966, respectively (Fig. 3a, b). This is within the accuracy range of 0.77–0.97 for commercial point-of-care testing coagulometers33. Bland-Altman analysis demonstrated a bias error of 1.617 s for PT, with 16 of 280 measurements samples falling outside the 95% agreement limits (Fig. 3c). Similar analysis for INR showed a bias error of −0.003, with 17 of 280 samples falling outside the 95% agreement limits (Fig. 3d). We note that increased variance at higher INRs may be due to variability in the international sensitivity index (ISI) of the tissue factor used by the laboratory versus our smartphone34. To evaluate test–retest reliability, each plasma sample was tested twice. The intra-assay coefficient of variation (CV) between duplicate measurements of 140 plasma samples was 3.62% for PT and 6.14% for INR, which is within the range of 1.4–8.4% found using commercial point-of-care testing coagulometers33.

For 100 of these plasma samples, we also conducted a manual tilt-tube test in parallel with the smartphone system. For these tests, 50 μl of plasma and 100 μl of activator were used; larger amount of plasma enabled more consistent testing with manual readings. The container was tilted back and forth until a clot formed. Clot times were noted by eye and recorded using a stop watch. Head-to-head testing demonstrated a PT/INR correlation between the manual test and the ground truth of R = 0.960 for both PT/INR, which is similar to the correlation obtained by our smartphone system (Supplementary Fig. 1a, b). The Bland-Altman analysis for PT showed a bias of −1.140 s, with five of 100 measurements falling outside the 95% agreement limits (Supplementary Fig. 1c). Similar analysis for INR showed a bias of 0.098, with six of 100 measurements falling outside these limits (Supplementary Fig. 1d).

### Coagulopathy testing

We also evaluated our system on an additional 79 plasma samples collected from two sites, from patients with a known coagulopathy. Specifically, we obtained samples across a broad range of coagulopathic causes including patients who had disseminated intravascular coagulation (DIC) (n = 8), liver disease (n = 19), trauma (n = 8), or conditions requiring anticoagulation such as extracorporeal membrane oxygenation (ECMO) (n = 6) or who were on heparin (n = 17) or warfarin (n = 13) to treat a medical condition (Table 1). The samples were obtained from both UWMC and Harborview Medical Center (HMC), and included samples from trauma patients in the emergency department undergoing blood transfusion. The PT/INR values of the plasma samples ranged from 12.6 to 67.2 s and 1–7.6, respectively, with a mean value of 22.2 s and 2.0 and a median value of 19.2 s and 1.6 (Supplementary Fig. 2b). We note that the mean PT/INR for these patients is double that of a normal INR of 1.0. Samples were collected and tested using the same procedure as the first evaluation on plasma samples.

The inter-class correlation coefficient between the smartphone system and the clinical-grade coagulation analyzer was R = 0.974 for both PT/INR. For samples with an elevated INR > 1.2, the correlation coefficient within each of the coagulopathy categories ranged from 0.890 to 0.977. Bland-Altman analysis demonstrated a bias error of −1.865 s for PT, with six of 158 measurements samples falling outside the 95% agreement limits (Fig. 4c). Similar analysis for INR showed a bias error of 0.060, with nine of 158 samples falling outside the 95% agreement limits (Fig. 4d).

On subgroup analysis, patients who were on heparin therapy did not demonstrate a substantially elevated PT/INR with a mean of 13.9 s and 1.1 respectively. Heparin affects coagulation of the intrinsic pathway, while PT/INR assesses the extrinsic pathway so this is an expected negative control result. Warfarin does directly affect the extrinsic pathway and this group had the highest average PT/INR of the different conditions studied with a mean of 36.2 s and 3.6 respectively. Individuals with an INR > 4.5 are at a nearly six fold increased risk of a bleeding event35. Of the four patients with an INR > 4.5, our system had an INR error of 14% compared to laboratory measurements. We included patients who quickly developed coagulopathy after a traumatic event as well as those with longstanding liver disease or on anticoagulation therapy. In all tested cases, the PT/INR assessment tools were well correlated, indicating that this device can be used with a variety of coagulopathies.

### Clinical testing on whole blood samples

We also evaluated the performance of our smartphone system on 80 anonymized samples of whole blood (blue top, 3 mL collected in sodium citrate) tested against the results of a commercial point-of-care PT/INR test meter (Coag-Sense, CoaguSense Inc.). In order to test the device capabilities across a range of coagulopathic conditions and PT/INR values, samples associated with particular diagnosis or from a particular clinic were preferentially obtained (the anticoagulation clinic and emergency department at Harboview Medical Center) with elevated PT/INR on laboratory testing, no other information about the sample than the patient coagulopathy was obtained. 30 of 80 samples were collected and tested on our smartphone system within 4 hours of being drawn from patients; twenty-two of the samples were tested within 4–12 h of blood draw; the remaining samples were refrigerated and tested more than 12 h later. The whole blood samples were collected under the same conditions as the plasma ones. PT/INR from the commercial test meter ranged from 13.5 to 39.0 s and 1.1–3.6, respectively, with a mean of 21.7 s and 1.9 and median of 21.1 s and 1.8 (Supplementary Fig. 2c).

The same amount of whole blood (10 μl) and thromboplastin activator (20 μl) were used for testing with the smartphone system and the commercial PT/INR meter. With the commercial meter, the whole blood and activator were each added to the test strip in quick succession as soon as the measurement started. The same thromboplastin activator was used for the commercial PT/INR meter and the smartphone system. As before, since each whole blood sample exceeded one milliliter, we tested each twice with the smartphone to evaluate test–retest performance.

The inter-class correlation coefficient was computed between the smartphone system and the commercial PT/INR meter (Fig. 5a, b). Across the 160 measurements of PT/INR, correlation coefficients were R = 0.936 and R = 0.933, respectively. These are within the accuracy range obtained by commercial point-of-care testing coagulometers33. Bland-Altman analysis for PT showed a bias of −0.843 s, with nine of 160 measurements falling outside the 95% limits (Fig. 5c). The bias for INR was 0.007, with ten of 160 measurements falling outside these limits (Fig. 5d).

We also evaluated the test–retest performance for whole blood testing. The intra-assay CV between the duplicate measurements was 5.39% for both PT/INR, which again is within the precision range obtained from commercial point-of-care testing coagulometers33 (Supplementary Table 2).

We evaluated the consistency of plasma and whole blood testing in our system by measuring a low and high PT/INR sample ten times in a row (Table 2). The low PT/INR (14.3 s and 1.1) plasma sample had a CV of 6.62% and 11.39%, respectively, while the high PT/INR (27.3 s and 2.5) sample had a CV of 4.52% and 7.76%, respectively. The low PT/INR (15.4 s and 1.3) whole blood sample had a CV of 8.06% for both PT/INR, while the high PT/INR (33.2 s and 3.0) sample had a CV of 6.64% for both PT/INR.

### Benchmark testing

Finally, we present benchmark testing across several design conditions. Testing was performed on plasma samples that were still able to clot more than 12 h after collection. Each scenario was tested three times on a single plasma sample, and the mean particle motion curve was plotted, with a shaded region representing one standard deviation from the mean. We omit the high amplitude spike at the start of the motion curve, which captures the activator addition into the solution, to more clearly identify the amplitude differences towards the end of the measurement. The motion curves are smoothed and cropped to show the area around the knee of the curve for visualization purposes.

We first considered the effect of vibration strength on particle motion (Fig. 6a). We varied vibration strength across a range of amplitudes from the Samsung Galaxy S9 smartphone’s built-in vibration motor and measured vibration strength at the cup holder with an accelerometer (Bosch Sensortec BMA400). For different vibration strengths, we computed the Euclidean norm from the x, y and z axes of the accelerometer and averaged them over a period of five seconds. The copper particle did not move as much at vibrations of 1.05 g compared to higher vibration values; however, for the full range of tested vibration values, the particle moved freely enough that the algorithm could detect when it stopped moving. Although the magnitude of particle movement did not change significantly for vibration strengths above 2.05 g in this benchmark, a vibration strength of 3.24 g was selected for our main evaluations: more viscous plasma samples required a higher vibration strength for the particle to move freely. Though the phone can produce vibrations of up to 3.77 g, we found that at those vibration levels, the particle could escape from the top of the container and small droplets of plasma spilled during measurement.

Second, we evaluated how different particle materials affected system performance (Fig. 6b). We found that small particles of copper, iron and foam could vibrate freely and thus were detectable. The copper and iron particles sank below the plasma surface and rattled around at the bottom of the container. The foam particle floated on the surface of the plasma. Although foam and iron particles produced sufficient motion in the fluid, we did not select them for the main evaluations since they were harder to reuse for subsequent samples. The plastic and cork particles both sank to the bottom of the container but still exhibited a small amount of motion that was detectable by the camera and algorithm. However, the amplitude differences between the vibratory and stationary states were small enough to be challenging to track across a larger number of measurements and samples.

Third, we evaluated how our system works with other smartphone models (Fig. 6c). The Samsung Galaxy S8 had a similar geometry to the Samsung Galaxy S9 used in the main evaluations, and produced comparable motion curves when the attachment was coupled to the top of the phone. The iPhone’s vibration motor had a lower strength, and the plastic attachment was placed in more direct contact with the right hand side of the phone where vibration amplitude was highest. While the Google Pixel’s vibration motor was located on the bottom of the phone, the motor was strong enough to produce comparable motion curves.

Fourth, we examined how ambient illumination affected the system (Fig. 6d). When the ambient incident light level was 0 lux with no external light sources, the smartphone’s flash was used to illuminate the cup and particle. For higher levels, the flash was turned off. At 0 lux, the flash reflected off the plasma and continued to move even after the plasma has coagulated. This results in high frequency motion seen throughout the motion curve. However, the particle’s transition to a stationary state could still be readily identified. At other illuminance levels, the camera’s ISO was increased and shutter speed decreased so the particle would be visible (see Materials and methods). The motion curves, and PT/INR at these illuminance levels were comparable and the particle’s stopping points could be identified.

Fifth, we examined the effects of different plasma volumes from 10 to 50 μl (Fig. 6e, f). In each case, the amount of thromboplastin added to the mixture was increased proportionately. We note that 10 μl represents the amount of whole blood (approximately one drop) required for use by the commercial point-of-care PT/INR meter, and 50 μl represents the amount of plasma used by the clinical-grade coagulation analyzer. Volumes of 10 and 20 μl were tested in the 4 mm diameter plastic tube containers used in our main evaluations. Larger volumes of plasma were tested in a larger 8 mm diameter plastic cup to accommodate the increased volume of fluid. The particle’s movement was more subtle at these higher volumes, but this can be attributed to the choice of container, which may have dampened vibrations. Across all volumes, measures of PT/INR were comparable, and the particle’s motion and transition to a stationary state was detectable by the algorithm.

Sixth, we evaluate the effect of placing the smartphone system on a variety of different surface materials (Supplementary Fig. 3). We chose both soft materials including foam and cloth, as well as hard materials that could form a tabletop surface including wood, metal and glass. We note that while the bottom half of the smartphone rests on the surface, the top half containing the vibration motor and attachment hangs over the edge of the surface to accommodate the shape of the attachment. We find that across all tested materials, the motion curves were similar and there was no significant change in PT/INR values. This shows that the smartphone vibration motor is strong enough to cause the particle to vibrate even when it is placed on common soft materials.

Seventh, we measured the effect of temperature and humidity on PT/INR for a single plasma sample on our smartphone system in a laboratory incubator (IVYX Scientific 5L Incubator). The incubator was used to control the ambient temperature around the smartphone and attachment to temperature values within the range of 18–32 °C). This range was selected to match the normal temperature conditions of use for the Coag-Sense POCT meter36. The relative humidity of the incubator ranged from 44–59% in this evaluation as measured by a hygrometer (Thermometer World). Across this temperature and humidity range, average PT/INR stayed relatively constant ranging from 12.9 to 13.6 s and 0.7–0.8 respectively (Supplementary Table 1). We note that this range is within the standard deviation of precision testing of a single low PT/INR plasma sample which is 1.1 s and 0.1 respectively (Table 2). As PT/INR testing typically takes less than a minute, ambient temperature and humidity, within normal limits, do not have a significant effect on system performance.

Finally, we examined the effect of diluting a whole blood sample with normal saline on PT/INR measurements from our smartphone system and a point-of-care PT/INR test meter. The hemoglobin levels of each dilution were also measured using a commercial hemoglobin meter (Mission Plus Hemoglobin Meter, Acon Laboratories Inc.). Dilution levels were selected such that the hemoglobin levels of the dilutions covered the range of 8.5–18.7 g/dL, which is the range within which PT/INR testing is to be performed in patients, as recommended by our institution37. We note that these dilution levels also affect the opacity of the blood sample. Supplementary Fig. 4 shows the dilution range from 0 to 50%, which covers hemoglobin concentrations in the range of 5.4–18.6 g/dL. As the dilution level increases from 0% to 50%, PT/INR values from our system and the POCT meter increase. PT/INR increased from 19.6 s (1.7) to 29.3 s (2.6) for the POCT meter and 18.5 s (1.5) to 32.2 s (2.7) for the smartphone system. This is because increased dilution levels dilute the coagulation factors that are activated as part of a PT assay, resulting in a longer time for a clot to form.

## Discussion

Our smartphone-based micro-mechanical clot detection system demonstrates strong correlation with laboratory and point of care PT/INR tests for both plasma and whole blood. 279 of 280 (99.6%) plasma measurements and 100 of 100 whole blood measurements fall within the allowable differences for INR testing, greater than the 90% threshold set by the International Organization of Standardization for this type of device38.

A key advantage of leveraging smartphone hardware for medical purposes is that custom electronics and hardware do not have to be designed, which lowers the development costs typically required to obtain regulatory approval. Specifically, under the FDA’s Mobile Medical Applications (MMA)39,40 and Software as a Medical Device (SaMD)41,42,43 guidance, the agency does not regulate the smartphone hardware, and only regulates custom software functions. The FDA has cleared or approved commercially available MMAs that use sensors like microphones and cameras to perform medical diagnostics44.

### Statistical analysis

Inter-class class correlation was calculated using Pearson’s correlation coefficient for correlation plots. The bias error (mean error) and 95% limits of agreement (LOA) were computed for the Bland-Altman plots. LOA was computed as 1.96 times the standard deviation of the error. For precision testing, CV was computed as the standard deviation divided by the mean of samples. For duplicate testing, intra-assay CV was computed as in ref. 65 where individual CVs were computed for each duplicate sample, and the mean CV across all samples was reported as the overall intra-assay CV. For benchmark testing, where repeated testing was performed under the same experimental conditions, PT/INR results were reported as mean ± standard deviation. Statistical analysis was performed using MATLAB, and the figures were generated using the Python matplotlib and seaborn library.

### Reporting summary

Further information on research design is available in the Nature Research Reporting Summary linked to this article.