Testing Near-Field Radio-Frequency Sensors to Predict Respiratory Distress in Patients Living With Dementia: A Pilot Feasibility and Acceptability Study

Abstract

Background

People living with dementia (PLWD) with advanced illness are prone to respiratory distress yet often cannot self-report dyspnea, delaying recognition and treatment. Near-field radio-frequency (NFRF) sensors offer touchless, covert cardiopulmonary monitoring that may be better tolerated than tethered devices.

Objectives

To assess the feasibility and acceptability of NFRF bed sensor for home monitoring of PLWD and to estimate machine-learning (ML) performance for detecting respiratory distress.

Methods

In a 48-hour pilot study, PLWD were recruited from a geriatrics practice. A lab-designed NFRF bed sensor recorded cardiopulmonary waveforms. Recorded video enabled minute-level Respiratory Distress Observation Scale scoring as reference. Feasibility outcomes included adverse events, acceptability, and percentage of usable data. ML classifiers (eg, random forest, k-nearest neighbors) were evaluated using 5-fold cross-validation, and class imbalance was addressed through data augmentation.

Results

Ten patient–legally authorized representative dyads were enrolled. No adverse events were reported, and no participants intentionally removed the sensor. Usable data averaged 52% (range 34-68%). Caregivers reported minimal burden and no patient distress. With augmented data, the random forest performed best, achieving 74.6% sensitivity and 95.5% specificity in detecting RDOS scores.

Conclusions

NFRF bed sensors were feasible and acceptable to implement in the home setting with PLWD, with promising ML-based detection of respiratory distress. Larger, longer studies with a broader range of RDOS severity are needed to validate performance and refine deployment. As this technology develops and matures, it could provide a method for non-invasive continuous monitoring to detect respiratory distress in PLWD in palliative care settings.

Keywords

dementia sensor technology respiratory distress palliative care caregivers

Introduction

Respiratory distress is a clinical manifestation of dyspnea often triggered by a pulmonary or cardiac cause (eg, pneumonia; heart failure), but it is also not uncommon to see at the end of life.¹ Whether in the community, long-term care, or hospice setting, patients living with dementia (PLWD) with advanced illness are at a heightened risk of developing respiratory distress due to their frailty¹ and risk of aspiration² – which can result in pneumonia and subsequent distress. Furthermore, PLWD with advanced illness often cannot communicate their needs to their caregivers and healthcare providers, which can lead to delayed detection, leading to unnecessary suffering and poor patient outcomes.^3–5

Technologies that continuously track and monitor biomedical data (eg, cardiopulmonary dynamics, body movement) have the potential to improve the delivery of medical and palliative care for this cohort of patients. For example, if a PLWD is in the advanced stages of their illness and is receiving home hospice care, detecting episodes of tachypnea might help guide medications to ease distress. However, present medical devices, such as polysomnography (PSG), respiratory belts, and photoplethysmography (ie, pulse oximeter), that estimate respiratory and/or heartbeat patterns continuously, are not typically implemented for long-term use, given that these devices⁶ require patients to be tethered to machines, wires, electrodes, or skin-contact clamps. However, near-field radio-frequency (NFRF) sensor technology has properties that enable the capture of biomedical data in a non-invasive, covert manner.^7,8 Their specific characteristics (eg, covert to patients, touchless, high sensitivity, robust against motion, and high multiplexing capabilities) make them advantageous for PLWD.^9–13 Furthermore, NFRF sensors can be designed in various forms, such as wearable devices, embedded in clothing, or integrated into beds. Coupled with machine learning (ML) methods, NFRF sensors have the potential to play an important role in continuously collecting cardiopulmonary measures and assisting healthcare providers in detecting respiratory distress, thereby leading to better symptom management and care for PLWD.

We have pilot tested NFRF sensors in patients with sleep apnea,¹⁴ COPD,¹⁵ and COVID-19,¹⁶ as well as healthy volunteers.⁸ In these studies, our team was able to detect and predict apneic events in patients undergoing sleep study, and showed the feasibility of a wearable sensor for COPD patients. However, no study has yet examined its use in PLWD. Therefore, we conducted a pilot study to assess whether NFRF sensors could be feasible and acceptable to monitor PLWD in the home setting. The objectives of this study were to (1) examine the feasibility and acceptability of implementing NFRF sensors in PLWD, and (2) calculate classification performance metrics (eg, sensitivity, specificity, positive predictive value, accuracy) using ML methods to detect levels of respiratory distress.

Methods

We conducted a pilot study to assess the feasibility and acceptability of using NFRF sensors in PLWD and applying ML methods to detect respiratory distress. This study was approved by Weill Cornell Medicine's Instituational Review Board (protocol # 22-03024639).

Participants

Recruitment of participants was conducted at [Blinded for Review]’s Center on Aging, an academic clinical practice consisting of geriatricians who provide care to older adults in New York City. PLWD were identified from electronic medical records and subsequently screened by their primary care provider on the appropriateness of enrolling the patient. Inclusion criteria for patients were (1) age ≥65 years old, (2) a diagnosis of dementia based on ICD-10 codes,¹⁷ and (3) a CDR-SB score ≥16¹⁸ (indicating a patient has severe dementia) as reported by the referring geriatrician. Exclusion criteria included (1) a patient with a pacemaker or implanted defibrillator (aimed to eliminate any small risk of electromagnetic interference with medical devices), and (2) a patient with known active pneumonia. We sought consent from the patient’s legally authorized representative (LAR). LAR inclusion criteria include age 18 years or older and English speaking.

Study Protocol

We contacted the LAR of the PLWD, and those who expressed interest were then scheduled for an in-person meeting at the patient’s residence. After obtaining written consent, we fitted the NFRF sensor, which we designed to be placed under the patient’s bed sheet (Figure 1). The decision to design such a covert sensor was based on avoiding any tethers or wires that could obstruct the patient’s caregiver routines or be accidentally removed by the patient. Figure 2 shows the depiction of the setup of our lab-designed NFRF bed sensor prototype.^9,14 The NFRF bed sensor was in place for 48 hours, and the LARs were informed to have the patients go about their usual routines and activities.

Study Measures

Patient demographic data were collected from the electronic medical record (EMR). We also confirmed demographic data from the patient’s LAR at the time of consent and conducted an exit interview with the LAR to understand feasibility and acceptability.

Our primary objective of testing feasibility and acceptability was measured by documenting (1) the percentage of potential participants who declined to participate, (2) whether the sensor was removed before study completion, (3) documented adverse events during sensor deployment, (4) a list of questions about device acceptability that was administered to the patient’s LAR at the end of the study, and (5) calculating the percentage of usable sensor data.

To develop our ML model, we needed to cross-validate the data collected from the bed sensor with clinical signs of respiratory distress. To achieve this, we set up a small video camera to record patients while they were in bed, and the bed sensor was in place. For each minute the patient was visible on the video recording, we calculated the patient’s respiratory distress using the Respiratory Distress Observation Scale (RDOS).¹⁹ For each minute for which we calculated an RDOS score, we defined it as an epoch. RDOS is a validated scale to identify respiratory distress in patients who cannot self-report dyspnea.²⁰ It is an 8-item scale that measures the severity of respiratory distress and provides a total score ranging from 0 to 16. A cutoff of 3 or greater on the RDOS indicates respiratory distress. Research staff who reviewed the videos were trained by a clinician (VP) to rate the RDOS. A clinician (VP) reviewed a subset of the data to ensure annotation accuracy with minimal operator bias.

Statistical Analysis

Descriptive statistics were used to describe feasibility and acceptability outcomes. Sensor data were pre-processed, down-sampled, and filtered to reduce high-frequency noises (eg, artifacts) that can degrade accuracy in the feature extraction phase. We then counted the number of epochs for our dataset, categorized by RDOS score.

ML algorithms were developed to detect respiratory distress using the sensor output waveform and annotated RDOS scores, including random forests, k-nearest neighbors (KNN), neural networks (NN), and decision trees.^21,22 These methods learn patterns in the data that distinguish between outcome groups and then assign each new case to the most likely group. For our analysis, we used a standard 5-fold cross-validation method, dividing our dataset into 5 parts and using 4 parts for training and 1 part for testing. All models are completely reset when the fold choices are changed to ensure the test folds remain unseen. For each ML classifier, we calculated the sensitivity, specificity, positive predictive value (PPV), accuracy, and AUC (Area Under the Curve) ROC (Receiver Operating Characteristics).

Given the medical stability of the participants we recruited, we encountered an imbalance in RDOS scores, with a greater number of periods in which patients had lower RDOS scores. To reduce selection bias and achieve better balance among the classes, we applied data augmentation methods.²³ Data augmentation is a commonly used approach in medical and biomedical ML studies when rare but clinically important events are underrepresented. For our dataset, we applied augmentation on the less frequent classes (ie, higher RDOS scores) to increase their representation while preserving the physiological characteristics of the original signals. This strategy helps improve model stability and reduces the risk that the analysis is driven primarily by the most common observations, leading to more balanced and clinically meaningful performance across RDOS severity levels.

Results

We identified 221 patients with a diagnosis of dementia from the EMR. Forty-two met the CDR-SB score of ≥16 and were contacted. 18 LARs declined to participate, 4 were interested but did not meet study criteria, and 10 were unreachable. We enrolled 10 patient/LAR dyads (26% recruitment rate). Table 1 shows the demographic data for patients and their LARs.

Table 1.

Demographic Data

Characteristics	Patient (N, %)	LAR (N, %)
Age (mean ± SD)	89.4 ± 7.65	71.1 ± 11.01
70-79	2 (20)	-
80-89	1 (10)	-
90-99	7 (70)	-
Gender
Female	7 (70)	7 (70)
Male	3 (30)	3 (30)
Ethnicity
Hispanic, Latino/a, or Spanish origin	1 (10)	1 (10)
Not Hispanic, Latino/a, or Spanish origin	7 (70)	9 (90)
Unknown	2 (2)	0 (0)
Race
White	6 (60)	8 (80)
Asian	1 (10)	1 (10)
Other	1 (10)	1 (10)
Unknown	2 (20)	0 (0)
Cardiac disease	1 (10)	-
Pulmonary disease	1 (10)	-
Highest education completed
Graduate degree	-	6 (60)
College degree	-	2 (20)
Some college	-	2 (20)
Relationship to patient
Child	-	6 (60)
Spouse/partner	-	4 (40)

Feasibility and Acceptability Data

We were able to place the NFRF bed sensor for all participants at their homes. All sensors were collected back after 48 hours. By removing noisy epochs and accounting for time the patient was off the bed, the percentage of usable data ranged from 34 % to 68%, with a mean of 52%. No participants (or caregivers) purposefully removed the sensor during the study period. For one participant, the NFRF sensor cables were accidentally dislodged from the connecting laptop by the patient’s Hoyer lift.

In our exit interviews with the LARs, 90% (n = 9) reported no concerns about using the NFRF sensor to monitor patients’ biomedical data. All LARs reported that the sensor did not cause any distress for the patient over the course of the 48-hour period. Only one LAR reported a potential barrier in using the sensor, which was that the wires connected to the sensor got in the way of the patient’s Hoyer lift. However, nine out of the ten LARs saw the benefits of having information on patients’ breathing and detecting issues early on, since patients cannot express their discomfort themselves. No adverse events were reported during the study.

Machine Learning Analyses

The distribution of un-augmented epochs across RDOS scores ranged from 0 to 5, with 10 020 epochs for an RDOS of 0, 2084 epochs for an RDOS of 1, 1039 epochs for an RDOS of 2, 229 epochs for an RDOS of 3, 56 epochs for an RDOS of 4, and 2 epochs for an RDOS of 5. Applying data augmentation methods, we increased our epoch numbers to 2500 epochs for an RDOS of 0, 2500 epochs for an RDOS of 1, 2500 epochs for an RDOS of 2, 2500 epochs for an RDOS of 3, 2500 epochs for an RDOS of 4, and 340 epochs for an RDOS of 5.

Table 2 summarizes the classification performance metrics (eg, sensitivity, specificity, positive predictive value, accuracy, and AUC-ROC) of 4 ML classifiers when using un-augmented data. For our 5-fold CV, the best performing model, random forest, achieved a sensitivity of 32.6% and a specificity of 89.7%, whereas neural net had poorer performance. Similarly, when calculating the AUC-ROC measure, decision tree and random forest emerged as acceptable at distinguishing distress at a threshold of ≥2, while distinguishing distress at higher scores was poor.

Table 2.

Classification Performance Metrics Using Machine Learning Models – Un-augmented RDOS Scores

Model	Sensitivity (%)	Specificity (%)	Positive predictive value (%)	CV accuracy (%)	AUC ROC (threshold $\geq 2$ )	AUC ROC (threshold $\geq 3$ )	AUC ROC (threshold $\geq 4$ )
Decision tree	32.2	90	51.3	79.5	0.738	0.554	0.508
Random forest	32.6	89.7	61.4	80.34	0.725	0.574	0.526
Neural net	22.5	85.3	59.5	76.8	0.662	0.525	0.509
KNN	30.9	88.4	74.9	78.1	0.716	0.583	0.526

CV, Cross-validation; AUC ROC, Area Under the Curve Receiver Operating Characteristics.

Table 3 summarizes the classification performance metrics (eg, sensitivity, specificity, positive predictive value, accuracy, and AUC-ROC) of 4 ML classifiers when using augmented data. For our 5-fold CV, the best performing model, random forest, achieved a sensitivity of 74.6% and a specificity of 95.5%, showing better performance in detecting higher distress, whereas neural net had poorer performance. When calculating the AUC-ROC measure, random forest emerged as consistently good at distinguishing distress across possible decision thresholds (0.865, 0.899, and 0.907 for RDOS thresholds of ≥2, ≥3, and ≥4).

Table 3.

Classification Performance Metrics Using Machine Learning Models - Augmented RDOS Scores

Algorithm	Sensitivity (%)	Specificity (%)	Positive predictive value (%)	CV accuracy (%)	AUC ROC (threshold $\geq 2$ )	AUC ROC (threshold $\geq 3$ )	AUC ROC (threshold $\geq 4$ )
Decision tree	69.1	93.7	69.6	71.7	0.840	0.858	0.855
Random forest	74.6	95.5	75.2	74.8	0.865	0.899	0.907
Neural net	42.5	89.5	52.0	54.5	0.677	0.732	0.705
KNN	66.0	92.9	66.9	67.7	0.798	0.828	0.845

CV, Cross-validation; AUC ROC, Area Under the Curve Receiver Operating Characteristics.

Discussion

To our knowledge, this is the first study to pilot NFRF technology for PLWD. We found that implementing the bed sensor over a 48-hour period was feasible and acceptable for participants. Further, our ML models showed a promising detection of various RDOS scores using classifiers such as Random Forest, KNN, and NN.

Previous reports in the literature have been mainly conducted in laboratory¹⁴ and/or structured clinical settings.¹⁵ In this real-world scenario, we were able to better understand the potential challenges in implementing this technology for PLWD in the home setting. One challenge, from a feasibility perspective, was that our sensor required cables to be connected to a laptop to record data continuously. We stored the cables and computer underneath the patient’s bed, which did not interfere with the patient’s care except for one situation where the patient’s Hoyer lift dislodged one of the wires. If this technology is adopted more widely for patient care, it will require more patient-friendly designs that are easy to use and do not disrupt daily activities. By implementing our prototype sensor with 10 patients, we believe future improvements needed for broader long-term care adoption include (1) a wireless sensor design, (2) a design that does not increase the risk of pressure ulcers, especially since many of these patients are bedbound and at higher risk of developing bed sores, and (3) a design capable of relaying biomedical data in real-time for analysis and interpretation.

The other important challenge we wanted to assess was understanding the quality of data collected in real-world conditions. Unlike our previous studies, which were conducted in a lab or clinical setting, in which protocols are much more controlled and structured, we wanted to pilot an NFRF sensor in a cohort of patients undergoing normal day-to-day activities without operator participation. We found that usable data was variable. In this study, the deployment of the NFRF sensor was under “real world” conditions – some patients were sitting up in bed or reclined in different positions. As a result, we expected some degradation in data quality. We identified factors (eg, caregiver movement when performing caregiving duties for the patient in bed) that impacted the data quality. Future work will focus on identifying ways to reduce data degradation so that interpretable data can be more consistently captured from participants.

Given that participants were only recorded for 48 hours and were in their usual state of health, we found a lack of significant variation in RDOS scores among our patient cohort, with most patients scoring a RDOS of 2 or less. To improve the accuracy of our ML models for detecting various levels of respiratory distress, future studies will need to include patients with a wider range of RDOS scores and incorporate longer prospective data.

Lastly, when applying ML methods on augmented data, we found promising results, particularly with random forest models, which showed decent sensitivity (75%), good specificity (95%), and good AUC-ROC. While this is based on a small sample size and more data are required before implementation in clinical settings, we can draw on other tests used in clinical practice for comparison. For example, a measure such as N-terminal pro-B-type natriuretic peptide (pro-BNP) which is commonly used to assist in detecting a heart failure exacerbation has a sensitivity of 92% and specificity of 88%,²⁴ while a test like d-dimer, used to help determine workup for deep vein thrombosis, has been shown to have a sensitivity range of 75-95% and a specificity range of 45-83% based on the type of assay used.²⁵ For future studies, improving sensitivity to avoid false positives RDOS scores will be important for implementation in clinical settings. Further, we believe understanding the patterns of RDOS scores, such as frequency and duration, will be important in assisting providers with the next steps in clinical care for these patients. For example, if the algorithm detects a patient with an RDOS of 7, which indicates severe distress, that lasts for only a couple of minutes but then resolves, it might not be as clinically significant compared to a patient with an RDOS of 5, indicating moderate distress, but that lasts continuously for 6 hours.

There were limitations in our study. While this pilot study looked at the feasibility over a period of 48 hours, a longer prospective study is needed to provide further information on its practicality. We recruited patients and their LARs from one clinical site, which may not reflect the diverse experiences of PLWDs and their LARs (or caregivers) or other care settings (eg, nursing home, home hospice care). Therefore, further testing in various care settings is needed.

Conclusions

This study demonstrates the potential of NFRF sensors to collect biomedical data in PLWD. Because this was a pilot study, additional patient data are needed to understand the heterogeneity of respiratory distress in the PLWD population and to build on our results to improve detection metrics that can be eventually useful for clinical care. As this technology develops and matures, it could provide a method for non-invasive, continuous monitoring of PLWD and other patients with serious and terminal illnesses.

Footnotes

ORCID iD

Veerawat Phongtankuel

Ethical Considerations

This study was approved by the Institutional Review Board at [Blinded for Review] on August 22, 2022. This research was conducted ethically in accordance with the Declaration of Helsinki.

Consent to Participate

Written informed consent was obtained from a legally authorized representative prior to enrollment in the study.

Author Contributions

Study concept and design: VP, ECK, JK, BH, KG. Acquisition of subjects and/or data: KG, TP, MDT, SJ, EY, CF, MS. Analysis and interpretation of data: VP, KG, ECK, JK, BH, TBC. Drafting and revising the manuscript: VP, KG, ECK, TP, VL. Final approval of the manuscript: All authors.

Funding

The authors disclosed receipt of the following financial support for the research, authorship, and/or publication of this article: Research reported in this publication/presentation was supported by the National Institute on Aging of the National Institutes of Health under Award Number P30AG073105.

Declaration of Conflicting Interests

The authors declared no potential conflicts of interest with respect to the research, authorship, and/or publication of this article.

Disclaimer

The content is solely the responsibility of the authors and does not necessarily represent the official views of the National Institutes of Health.

References

Campbell

Kiernan

Strandmark

Yarandi

. Trajectory of dyspnea and respiratory distress among patients in the last month of life. J Palliat Med. 2018;21(2):194-199. doi:10.1089/jpm.2017.0265

Manabe

Teramoto

Tamiya

Okochi

Hizawa

. Risk factors for aspiration pneumonia in older adults. PLoS One. 2015;10(10):e0140060. doi:10.1371/journal.pone.0140060. In: Chalmers JD, ed.

Meehan

Fine

Krumholz

, et al. Quality of care, process, and outcomes in elderly patients with pneumonia. JAMA. 1997;278(23):2080-2084.

Houck

Bratzler

Nsa

Bartlett

. Timing of antibiotic administration and outcomes for medicare patients hospitalized with community-acquired pneumonia. Arch Intern Med. 2004;164(6):637-644. doi:10.1001/archinte.164.6.637

Lee

Nsa

Hausmann

LRM

, et al. Quality of care for elderly patients hospitalized for pneumonia in the United States, 2006 to 2010. JAMA Intern Med. 2014;174(11):1806-1814. doi:10.1001/jamainternmed.2014.4501

Fraden

. Handbook of Modern Sensors. New York: Springer; 2010, Vol. 3. doi:10.1007/978-1-4419-6466-3

Hui

Kan

. Monitoring vital signs over multiplexed radio by near-field coherent sensing. Nat Electron. 2017;1(1):74-78. doi:10.1038/s41928-017-0001-0

Zhang

Zhou

Conroy

, et al. Near-field radio sensing for biomedical, biological, and cyberphysical intelligence. IEEE Sens Lett. 2024;8(3):1-4. doi:10.1109/LSENS.2024.3356817

Zhang

Sharma

Zhou

Hui

Kan

. Furniture-integrated respiration sensors by notched transmission lines. IEEE Sens J. 2021;21(4):5303-5311. doi:10.1109/JSEN.2020.3028970

10.

Zhang

Sharma

Conroy

Phongtankuel

Kan

. Objective scoring of dyspnea by non-invasive RF respiratory sensors. IEEE Trans Biomed Eng. 2022;69(1):432-442. doi:10.1109/TBME.2021.3096462

11.

Will

Shi

Schellenberger

, et al. Radar-based heart sound detection. Sci Rep. 2018;8(1):11551. doi:10.1038/s41598-018-29984-5

12.

James

. Diagnostic applications of electromagnetic fields. In: Stone

, eds. The Review of Radio Science: 1990-1992 . USA: Oxford science publications. publ. for the International union of radio science by Oxford university press; 1993:771-778.

13.

Lubecke

Boric-Lubecke

Lin

. A review on recent advances in doppler radar sensors for noncontact healthcare monitoring. IEEE Trans Microw Theor Tech. 2013;61(5):2046-2060. doi:10.1109/TMTT.2013.2256924

14.

Zhang

Conroy

Krieger

Kan

. Detection and prediction of sleep disorders by covert bed-integrated RF sensors. IEEE Trans Biomed Eng. 2023;70(4):1208-1218. doi:10.1109/TBME.2022.3212619

15.

Zhang

Park

Krishnan

, et al. A feasibility study using radio-frequency sensors to collect respiratory metrics in patients with chronic obstructive pulmonary disease. Med Res Arch. 2024;12(1):4910. doi:10.18103/mra.v12i1.4910

16.

Zhang

Zhou

Conroy

, et al. Deduced respiratory scores on COVID-19 patients learning from exertion-induced dyspnea. Sensors. 2023;23(10):4733. doi:10.3390/s23104733

17.

Centers for Medicare & Medicaid

Services.

ICD 10 CM/PCS MS DRG Definitions Manual, version 37.0: Major Diagnostic Category 19 Mental Diseases & Disorders (Organic Disturbances and Intellectual Disability). Centers for Medicare & Medicaid Services; 2019. https://www.cms.gov/icd10m/version37-fullcode-cms/fullcode_cms/P0334.html. Accessed 1 August 2024.

18.

O’Bryant

Lacritz

Hall

, et al. Validation of the new interpretive guidelines for the clinical dementia rating scale sum of boxes score in the national Alzheimer’s coordinating center database. Arch Neurol. 2010;67(6):746-749. doi:10.1001/archneurol.2010.115

19.

Campbell

Templin

Walch

. A respiratory distress observation scale for patients unable to self-report dyspnea. J Palliat Med. 2010;13(3):285-290. doi:10.1089/jpm.2009.0229

20.

Campbell

. Psychometric testing of a respiratory distress observation scale. J Palliat Med. 2008;11(1):44-50. doi:10.1089/jpm.2007.0090

21.

Wen

Zhang

. Deep convolution neural network and autoencoders-based unsupervised feature learning of EEG signals. IEEE Access. 2018;6:25399-25410. doi:10.1109/ACCESS.2018.2833746

22.

Uddin

Khan

Hossain

Moni

. Comparing different supervised machine learning algorithms for disease prediction. BMC Med Inf Decis Making. 2019;19(1):281. doi:10.1186/s12911-019-1004-8

23.

Mumuni

. Data augmentation: a comprehensive survey of modern approaches. Array. 2022;16:100258. doi:10.1016/j.array.2022.100258

24.

Han

Cheng

, et al. Diagnostic accuracy of natriuretic peptides for heart failure in patients with pleural effusion: a systematic review and updated meta-analysis. PLoS One. 2015;10(8):e0134376. doi:10.1371/journal.pone.0134376. In: Lionetti V, ed.

25.

Weitz

Fredenburgh

Eikelboom

. A test in context: D-dimer. J Am Coll Cardiol. 2017;70(19):2411-2420. doi:10.1016/j.jacc.2017.09.024