Design and Performance of a Multisensor Heart Failure Monitoring Algorithm: Results From the Multisensor Monitoring in Congestive Heart Failure (MUSIC) Study

Journal of Cardiac Failure Vol. 18 No. 4 2012

Design and Performance of a Multisensor Heart Failure Monitoring Algorithm: Results From the Multisensor Monitoring in Congestive Heart Failure (MUSIC) Study INDER S. ANAND, MD, FRCP, DPhil (Oxon),1 W.H. WILSON TANG, MD,2 BARRY H. GREENBERG, MD,3 NIRANJAN CHAKRAVARTHY, PhD,4 IMAD LIBBUS, PhD,4 AND RODOLPHE P. KATRA, PhD,4 ON BEHALF OF THE MUSIC INVESTIGATORS Minneapolis and St. Paul, Minnesota; Cleveland, Ohio; and San Diego, California

ABSTRACT Background: Remote monitoring of heart failure (HF) patients may help in the early detection of acute decompensation before the onset of symptoms, providing the opportunity for early intervention to reduce HF-related hospitalizations, improve outcomes, and lower costs. Methods and Results: MUSIC is a multicenter nonrandomized study designed to develop and validate an algorithm for prediction of impending acute HF decompensation with the use of physiologic signals obtained from an external device adhered to the chest. A total of 543 HF patients (206 development, 337 validation) with ejection fraction #40% and a recent HF admission were enrolled. Patients were remotely monitored for 90 days using a multisensor device. Accounting for device failure and patient withdrawal, 314 patients (114 development, 200 validation) were included in the analysis. Development patient data were used to develop a multiparameter HF detection algorithm. Algorithm performance in the development cohort had 65% sensitivity, 90% specificity, and a false positive rate of 0.7 per patient-year for detection of HF events. In the validation cohort, algorithm performance met the prespecified end points with 63% sensitivity, 92% specificity, and a false positive rate of 0.9 per patient-year. The overall rate of significant adverse skin response was 0.4%. Conclusion: Using an external multisensor monitoring system, an HF decompensation prediction algorithm was developed that met the prespecified performance end point. Further studies are required to determine whether the use of this system will improve patient outcomes. (J Cardiac Fail 2012;18:289e295) Key Words: Heart failure, remote monitoring, intrathoracic impedance, acute decompensated heart failure, fluid retention, HF detection algorithm.

Heart failure (HF) is a leading cause of mortality and morbidity. Despite advances in the management of HF, hospitalizations for acute decompensation remain high, reaching more than 1.1 million discharges annually and accounting for the bulk of the $37.2 billion spent on these patients in 2009.1 The most frequent cause of an HF-related hospital admission is fluid retention, and current HF

disease management guidelines recommend monitoring fluid status in patients with HF.2,3 Tracking weight gain as a surrogate for fluid retention has been commonly used to assess changes in HF status, but numerous studies have reported its lack of sensitivity, poor specificity, and high false positive rate.4e8 Implantable devices with special fluid monitoring features can detect pathophysiologic deteriorations in HF patients weeks before symptom onset, and they have better sensitivity and specificity than serial body weight measurements for predicting HF decompensation.7,9,10 However, the majority of HF patients, particularly those with HF and preserved ejection fraction (EF), are not candidates for such devices, which are also extremely expensive, prohibiting broad adoption. Nonimplantable wireless devices capable of continuously monitoring patient fluid status and other physiologic signals may play a role in detecting impending HF decompensation without being subject to some of the limitations of implantable devices.

From the 1Veterans Administration Medical Center, Minneapolis, Minnesota; 2Cleveland Clinic, Cleveland, Ohio; 3University of California, San Diego, California and 4Corventis, St. Paul, Minnesota. Manuscript received June 27, 2011; revised manuscript received December 13, 2011; revised manuscript accepted January 5, 2012. Reprint requests: Inder S. Anand, MD, FRCP, DPhil (Oxon), Director, Heart Failure Program, VA Medical Center 111C, Minneapolis, MN 55417. Tel: 612-467-3663; Fax: 612-970-5899. E-mail: anand001@umn. edu All decisions regarding this manuscript were made by a guest editor. See page 294 for disclosure information. 1071-9164/$ - see front matter Published by Elsevier Inc. doi:10.1016/j.cardfail.2012.01.009

289

290 Journal of Cardiac Failure Vol. 18 No. 4 April 2012 Although current impedance-based HF detection technology is superior to the use of weight gain for detection of HF decompensation, it lacks a high degree of specificity and prognostic power on an individual patient level.10,11 Therefore, improved monitoring devices are required to detect HF status changes on a predictive and personalized level. HF detection algorithms using multiple physiologic variables, each with its independent prognostic value, may result in a combined prognostic index of superior predictive and personalized value to an algorithm based on a single variable. The Multisensor Monitoring in Congestive Heart Failure (MUSIC) study was designed to develop and validate a personalized multiparameter algorithm to predict acute HF decompensation (ADHF) using a noninvasive multisensor adherent monitoring system. Methods MUSIC Study Design and Objective The design of the MUSIC study has been described in detail previously.12 Briefly, MUSIC was a multicenter nonrandomized study with 2 phases. In the MUSIC-Development phase, physiologic and clinical data were collected and used to develop a personalized multiparameter algorithm to detect impending HF decompensation. The algorithm was subsequently validated on a separate patient cohort in the MUSIC-Validation phase. The algorithm had a composite target performance objective of detecting ADHF events with $60% sensitivity and a false positive rate of #1.0 per patientyear and a safety objective of a #5% severe adverse event rate. The MUSIC study enrolled a total of 543 patients in New York Heart Association (NYHA) functional class III or IV, with left ventricular EF #40% and a history of hospitalization for HF of at least 1 within 60 days before enrollment or 2 in the 12 months before enrollment. Patients with known allergy or hypersensitivity to adhesives or hydrogel were excluded from the study, as were patients with an implantable device with an active minute ventilation sensor. The first 206 patients were assigned to the MUSICDevelopment phase, and the remaining 337 patients were assigned to the MUSIC-Validation phase. Although MUSIC-Development and MUSIC-Validation phase patients were enrolled as part of a single protocol, all data from MUSIC-Validation patients remained sequestered until algorithm development was complete. No patient was used for both algorithm development and algorithm validation. Study patients were monitored for 90 days with the MUSE system, a clinical prototype device capable of continuously recording physiologic variables such as bioimpedance, heart rate, respiration, activity, and posture.12 Study Procedures The adherent device was applied to the patient’s chest at enrollment and replaced weekly during the 90-day monitoring period. The MUSE monitoring system kit provided to the patients at the time of enrollment contained 13 adherent devices, 1 for each week, for a total of 13 weeks of monitoring. The first device was applied in the clinic at enrollment, and the patient was instructed on the procedure of application and removal of the patch device at weekly intervals for the 90-day duration of monitoring. During the follow-up period, patients were managed in the usual standard of care by the primary provider, who was not provided

with device data. Each study site made telephone follow-up calls at 1 week, 30 days, and 60 days after enrollment to document any HF-related adverse events or medication changes that may have occurred during the study. At study completion, patients returned the MUSE system kit and completed a brief survey about their experience with the MUSE system. HF Detection and Classification Patients were classified into 2 clinical groups: patients with no ADHF events during enrollment, classified as ‘‘condition absent,’’ and patients with at least one ADHF event, classified as ‘‘condition present.’’ An ADHF event was defined as any of the following: 1. Any HF-related hospitalization, emergency room or urgent care visit that required administration of IV diuretics, inotropes, or ultrafiltration for fluid removal. 2. A change in diuretic (including aldosterone antagonists) directed by the health care provider that included one or more of the following: 1) a change in the prescribed diuretic type; 2) an increase in dose of an existing diuretic; or 3) the addition of another diuretic. 3. An ADHF event for which death was the outcome. Patient status signals generated by the MUSE system algorithm were classified as: 1. True positive: a positive patient status signal that lasted $24 hours and culminated in an ADHF event. 2. False positive: a positive patient status signal without a corresponding ADHF event, or a positive patient status signal that switched to negative without a corresponding ADHF event. 3. False negative: an ADHF event that was not preceded by a positive patient status signal in the preceding 24 hours. Repeated ADHF events that occurred O48 hours apart were considered to be separate events. In addition, if a patient was enrolled during an event, that event was not included in the analysis, and the alert status for that patient was not analyzed until 48 hours after the event had ended. 4. True negative: a negative patient status signal with no ADHF events for the duration of the study. Sensitivity was defined as the percentage of true positive events out of all ‘‘condition present’’ events. Specificity was defined as the percentage of all ‘‘condition absent’’ patients who had no false positive events. False positive rate was defined as the number of false positive patient status signals per patient-years of follow-up. Safety End Point The safety end point was defined as the percentage of patients who developed significant skin irritation, contact dermatitis, or skin damage related to the adherent device that resulted in physician recommendation to remove and discontinue use of the monitoring system.

Results Patient Demographics and Baseline Characteristics

The baseline characteristics of the 543 patients enrolled in the MUSIC study from 27 centers in the US, India, and Singapore are summarized in Table 1. The overall patient population was predominantly male. The baseline

Multisensor Monitoring in CHF

characteristics of the development and validation cohorts had similar body mass index, EF, blood pressure, heart rate, and incidence of hypertension and ischemic heart disease. There was a significant difference in NYHA functional class and use of background HF therapy. A total of 229 patients (92 development, 137 validation) were excluded from analysis for the following reasons: 127 patients (56%) had device failure; 59 patients (26%) withdrew consent after enrollment; 3 patients (1%) were withdrawn by the primary investigator; and 32 patients (14%) had contraindicating conditions. The baseline characteristics of the excluded patients were not significantly different from the remaining 314 patients (114 development, 200 validation). A comparison between enrolled and included patients in the validation phase is presented in Table 2. Out of all patients (n 5 314) included in the analysis, 62 had at least 1 ADHF event during enrollment, for a patient event rate of 19.7%. These 62 patients had a total of 115 events (54 hospitalizations or emergency room visits, 37 diuretic changes, and 24 deaths), for an overall event rate of 36.6% in the 314 patients. Safety

The device met the prespecified safety end point, with only 2 patients (0.4%), both in the development cohort, experiencing severe adverse events related to wearing the MUSE system. Algorithm Development

The signals from the MUSE system were collected continuously for 90 days and analyzed offline to develop the algorithm to predict heart failure decompensation events. Raw signals from the impedance, electrocardiographic, and accelerometer sensors were filtered and stored as discrete parameters that could be used either individually or in combination for ADHF detection algorithm development. The derived parameters included bioimpedance, heart rate, respiratory rate and volume, activity duration and Table 1. Baseline Demographics and Characteristics of Patients Enrolled in the MUSIC Study

Age (y, mean 6 SD) Sex (% male) History of ischemic heart disease (%) History of hypertension (%) NYHA functional class III (%)* BMI (kg/m2, mean 6 SD) Ejection fraction (%) Systolic BP (mm Hg, mean 6 SD) Heart rate (beats/min, mean 6 SD) Medication Use* ACEi/ARB (%) b-Blockers (%) Aldosterone antagonists (%) Diuretics (%)

Development (n 5 206)

Validation (n 5 337)

60 6 13 72 62 44 94 25 6 6 27 6 6 118 6 19 81 6 15

57 6 13 78 61 45 63 26 6 6 27 6 7 117 6 17 79 6 13

95 55 35 78

84 73 45 96

NYHA, New York Heart Association; BMI, body mass index; BP, blood pressure; ACEi, angiotensin-converting enzyme inhibitor; ARB, angiotensin receptor blocker. *P ! .05.



Anand et al

291

Table 2. Comparison of Baseline Characteristics of Patients in the Validation Group That Were Enrolled and Those Whose Data was Included in Analysis

Age (y) Sex (% male) BMI (kg/m2) Ischemic heart disease (%) Hypertension (%) NYHA functional class III (%) Ejection fraction (%) Systolic BP (mm Hg) Heart rate (beats/min) ACEi/ARB (%) b-Blockers (%) Diuretics (%) Aldosterone antagonists (%)

Enrolled (n 5 337)

Included (n 5 200)

P Value

57 6 13 78 26 6 6 61 45 63 27 6 7 117 6 17 79 6 13 84 73 96 45

59 6 12 78 25 6 5 65 43 61 27 6 7 118 6 18 79 6 13 85 70 96 45

ns ns ns ns ns ns ns ns ns ns ns ns ns

Abbreviations as in Table 1.

intensity, and body posture. These parameters were evaluated for correlation with documented ADHF events. MUSIC-Development patient data were used to develop a single-parameter algorithm. That algorithm was then applied to all of the MUSIC-Validation patients. Single-Parameter Algorithm Using Bioimpedance

The single-parameter algorithm was developed using a proprietary index of fluid status, which includes a continuous bioimpedance measurement normalized to the patient’s own baseline. The deviation from baseline is normalized and scaled; negative values indicate fluid accumulation, while positive values indicate fluid loss. Figure 1 shows the sensitivity, specificity, and false positive rate for ADHF event detection based on the single-parameter fluid index in a receiver operating characteristic (ROC) curve. At a fluid index threshold change of
1.8, the sensitivity for HF prediction was within our prespecified range, but corresponded to a false positive rate of 5.3 per patient-year. Therefore, the single-parameter algorithm was not adequate to meet the prespecified performance end point (sensitivity $60% with a false positive rate of #1.0 per patient-year). Multiparameter Algorithm

To address the shortcomings of a single-parameter algorithm, a proprietary personalized, tiered, multiparameter ADHF detection algorithm was developed. The algorithm was composed of 3 primary components: 1. A fluid index that measures normalized changes in bioimpedance from the patient’s baseline, expressed as standard deviation change from baseline status. 2. A breath index that serves as a measure of shallow breathing. This index also measures relative deviations from baseline breath index, expressed as standard deviation from normal status. 3. Personalization parameters that correct baseline impedance to patient characteristics such as age, sex, height, and weight.

292 Journal of Cardiac Failure Vol. 18 No. 4 April 2012 100

18

Sensitivity

80

16

70

14

60

12

50

10

False Positive Rate

40

8

30

6

20

4

10

2

0

Event / Patient-Year

0 -8

-6

-4

-2

0

2

4

6

8

Algorithm Performance

10

Fluid Index Threshold

The performance of the algorithm using the singleparameter algorithm and 3 different multiparameter algorithm permutations is presented in Table 3. In the development cohort, the single-parameter algorithm based on fluid index alone had a relatively high sensitivity (83%) but poor specificity and false positive rate (40% and 9.4 events per patient-year, respectively). The 2-parameter algorithms combining fluid index and breath index or fluid index and personalized fluid status resulted in improved specificity, false positive rate, and positive predictive value with a moderate reduction in sensitivity. However, using the 3 parameters in combination (fluid index, breath index, and personalized fluid status) yielded a robust algorithm with a sensitivity of 65%, a specificity of 90%, and a false positive rate of 0.7 events per patient-year, meeting the specified performance end points. The 3-parameter algorithm that was generated in the development phase was then applied to the validation cohort (Table 4). In the validation cohort, the algorithm achieved a sensitivity of 63%, specificity of 92%, and a false positive rate of 0.9 events per patient-year, also meeting the specified performance end points. The advance detection time preceding an ADHF event averaged 9.4 days in the development cohort and 11.5 days in the validation cohort. The advance detection time

Fig. 1. Single-parameter receiver operating characteristic curve, showing sensitivity, specificity, and false positive rate for acute heart failure decompensation event detection based on the fluid index. The crossover point (square) corresponds to a sensitivity and specificity of 68% with a false positive rate of O4 per patientyear. The optimal threshold (circle) yields a sensitivity of 83%, a specificity of 40%, and false positive rate of O9.4 per patientyear. The single-parameter algorithm was not adequate to meet the prespecified performance end point (sensitivity $60% and false positive rate of #1.0 per patient-year).

A typical ROC analysis is based on comparing a test’s sensitivity and specificity (or false positive rate) for different detection thresholds. Because the multiparameter ADHF detection algorithm was based on multiple indices (each with its own sensitivity and specificity at its corresponding thresholds), a combinatorial multiparameter ROC approach was used to determine the unique combination of thresholds from the algorithm. The algorithm performance was computed for all combinations of a wide range of fluid index thresholds, breath index thresholds, and personalized fluid status thresholds. The sensitivity, specificity, and false positive detection rate for the multiparameter sweep are shown in Figure 2. Panels A and B show the algorithm detection

100 80 60 40 20 0 0

B

Specificity

A

10

20

30

40

C

60

70

80

90

100

0

10

20

30

40

50

60

Sensitivity

10

20

30

40

10

20

30

40

D

Specificity

100 80 60 40 20 0

50

Sensitivity

12 10 8 6 4 2 0 0

False Positive Rate

-10

70

80

90

100

False Positive Rate

Percent

90

performance for the family of combinations of the 3 main indices’ thresholds. The outer envelopes of these pseudoROC plots (panels C and D) show the optimal specificity and false positive rate values at different sensitivities, with each data point corresponding to a different combination of threshold values. One of the combinations of thresholds (circled in Fig. 2D) yielded HF event detection that met the prespecified performance criteria (sensitivity $60% and false positive rate #1 per patient-year). As Panel D shows, the sensitivity can be improved at the expense of a higher false positive rate.

20

Specificity

12 10 8 6 4 2 0

0

50

60

70

80

90

100

50

60

70

80

90

100

Sensitivity

Sensitivity

Fig. 2. Multiparameter pseudoereceiver operating characteristic analysis. Panels A and B show the multiparameter sweep of different threshold combinations for sensitivity vs specificity and sensitivity vs false positive rate. Panels C and D show the outer envelope of the multiparameter sweep representing the best-performing combination for sensitivity vs specificity and sensitivity vs false positive rate. An optimal threshold combination, meeting the prespecified performance endpoint, is indicated (circle).

Multisensor Monitoring in CHF



Anand et al

293

Table 3. Performance Results for Algorithm Development Fluid Index

Fluid Index þ Breath Index

Fluid Index þ Personalization

Fluid Index þ Breath Index þ Personalization

83 40 9.4

70 77 2.3

71 74 3.2

65 90 0.7

Sensitivity (%) Specificity (%) False positive rate (events/patient-year)

varied widely, ranging from 2 days to O30 days. On average, the multiparameter and personalized algorithm had detection times that were not significantly different from the less specific algorithms. Discussion The MUSIC study was designed to develop and validate a HF detection algorithm using physiologic data recorded from a multisensor noninvasive skin-adherent monitoring system. We found that the personalized, tiered, multiparameter algorithm was capable of detecting HF worsening ahead of symptom onset and health care intervention, meeting the predefined end points with a sensitivity of 65% and false positive rate of 0.7 events per patient-year in the development cohort and a sensitivity of 63% and false positive rate of 0.9 events per patient-year in the validation cohort. The multisensor system also met the safety end point with a 0.4% severe adverse event rate related to prolonged wear of the adherent device. Multiparameter HF Detection Algorithm

The HF detection algorithm developed in this study was designed to address the unmet need for a noninvasive personalized HF detection and monitoring system with a high sensitivity and low false positive rate.10,11 To approach this problem, the development process began with the simplest design, a single-sensor fluid status measurement based on thoracic impedance to detect fluid accumulation in a manner similar to previously reported approaches.9 The overall performance of the single-sensor impedance-based approach was poor, with high sensitivity but mediocre specificity and a high false positive rate. The algorithm was then enhanced with data from additional physiologic sensors to improve the specificity of HF detection with lower false detection rates. The resulting proprietary algorithm relied on 3 central indices: a fluid index, a breath index, and a personalization adjustment. Table 4. Performance Results for Algorithm Validation

Sensitivity (%) Specificity (%) False positive rate (events/patient-year) Advance warning time (d)

Development Phase

Validation Phase

65 90 0.7 9.4 6 10.9

63 92 0.9 11.5 6 9.0

In the multiparameter algorithm, the overall performance depends on the thresholds used for each individual parameter. Therefore, to objectively calculate the optimal combination of thresholds for the ADHF detection multiparameter algorithm, a combinatorial sweep of a range of individual index thresholds was performed, resulting in a family of possible performance outcomes.13,14 The envelope of the resulting performance outcomes yields a pseudo-ROC curve that describes the performance of a multiparameter algorithm. The final threshold combination chosen for the algorithm was based on the specified performance end points. This proprietary multitiered, multiparameter, personalized ADHF detection algorithm met the prespecified performance end points in both the development cohort and the distinct validation cohort. Study Limitations

There are several limitations to this study. Out of the 543 enrolled patients in this study, 229 patients (42%) were excluded from the analysis. The exclusion rate was 45% in the MUSIC-Development phase and 41% in the MUSICValidation phase. Most of the exclusions were due to failure of the prototype device and for withdrawal of consent. However, there were no significant differences in baseline demographics between patients included or excluded from analysis, suggesting that the withdrawals did not bias the patient population. Although the demographics and other baseline characteristics of the patients enrolled in this study are typical of HF populations, the use of b-blockers and aldosterone antagonists was lower in the development than in the validation cohort, also reflected in the elevated baseline average heart rate.15e20 This was because, compared with the MUSICValidation cohort, more MUSIC-Development patients (88% vs 66%) were from outside the United States, where use of these drugs was not optimal. The geographical disparity in the use of these drugs also accounts for the higher event rates in the development cohort compared with the validation cohort (47% vs 30%). However, despite significant differences in the use of the standard of care medications in the MUSIC-Development and MUSIC-Validation groups, the HF detection algorithm met the performance end point in both populations, confirming its performance and applicability in diverse populations of heart failure patients. The performance of the MUSE clinical prototype was responsible for most of the 30% patient data loss (71% of all excluded patients, including device failure and device-

294 Journal of Cardiac Failure Vol. 18 No. 4 April 2012 related early withdrawal), for a variety of reasons, including device electronics failure, communications connection failure, and lack of real-time data monitoring. These factors were addressed as the investigational prototype was refined into a commercial clinical monitoring system, and subsequent clinical study experience has demonstrated patient data loss (including device-related and nondevice-related withdrawals) of !4%.

3.

Clinical Implications

The present study describes the development and validation of an ADHF detection algorithm with high sensitivity and low false detection rate that could be used to monitor HF status in patients without requiring an implantable device. Such a system could provide physicians with a tool for closely monitoring HF patient status, potentially reducing HF hospitalizations and lowering cost of care by promoting early intervention. The monitoring system may prove to be useful in the management of patients being discharged after an ADHF event to prevent 30- to 60-day rehospitalization. Studies to test this hypothesis are being planned. In the present study, the patients were monitored for 90 days. In commercial use, it is likely that monitoring will occur for only up to 30 days in an attempt to minimize 30-day readmission rate. Whether this monitoring approach will reduce 30-day or 60-day rehospitalization rates, and whether such an approach is cost-effective compared with current clinical judgment, will be tested in a subsequent study. With a 90-day event rate of 36.6% and a false positive rate of 0.9/patient-year, one would expect approximately 1.5 true events for each false positive. Because the cost of a true event (an HF hospitalization) is likely to be much greater than the cost of a false positive (an unnecessary clinic visit), such a ratio may be considered to be cost-effective. This ratio can be improved by restricting monitoring to particularly high-risk patients or other patients with higher event rates (such as immediately after discharge from an acute HF hospitalization).

4.

5.

6.

7.

8.

9.

10.

11.

12.

Disclosures This study was sponsored by Corventis. Drs Anand and Greenberg serve on the advisory board of Corventis but do not have equity stakes in the company. Drs Chakravarthy, Libbus, and Katra are current or former employees of Corventis. Dr Tang reports no potential conflict of interest.

13. 14.

15.

References 16. 1. Lloyd-Jones D, Adams R, Carnethon M, de Simone G, Ferguson TB, Flegal K, et al. Heart disease and stroke statisticsd2009 update: a report from the American Heart Association Statistics Committee and Stroke Statistics Subcommittee. Circulation 2009;119:480e6. 2. Swedberg K, Cleland J, Dargie H, Drexler H, Follath F, Komajda M, et al. Guidelines for the diagnosis and treatment of chronic heart failure: executive summary (update 2005): the Task Force for the

17.

18.

Diagnosis and Treatment of Chronic Heart Failure of the European Society of Cardiology. Eur Heart J 2005;26:1115e40. Hunt SA, Abraham WT, Chin MH, Feldman AM, Francis GS, Ganiats TG, et al. ACC/AHA 2005 guideline update for the diagnosis and management of chronic heart failure in the adult: a report of the American College of Cardiology/American Heart Association Task Force on Practice Guidelines (Writing Committee to Update the 2001 Guidelines for the Evaluation and Management of Heart Failure): developed in collaboration with the American College of Chest Physicians and the International Society for Heart and Lung Transplantation: endorsed by the Heart Rhythm Society. Circulation 2005; 112:e154e235. Zhang J, Goode KM, Cuddihy PE, Cleland JG. Predicting hospitalization due to worsening heart failure using daily weight measurement: analysis of the Trans-European NetworkeHome-Care Management System (TEN-HMS) study. Eur J Heart Fail 2009;11:420e7. Lewin J, Ledwidge M, O’Loughlin C, McNally C, McDonald K. Clinical deterioration in established heart failure: what is the value of BNP and weight gain in aiding diagnosis? Eur J Heart Fail 2005;7:953e7. Goldberg LR, Piette JD, Walsh MN, Frank TA, Jaski BE, Smith AL, et al. Randomized trial of a daily electronic home monitoring system in patients with advanced heart failure: the Weight Monitoring in Heart Failure (WHARF) trial. Am Heart J 2003;146:705e12. Abraham WT, Compton S, Foreman B, Haas G, Canby RC, Fishel R, et al. Superior performance of intrathoracic impedance-derived fluid index versus daily weight monitoring in heart failure patients: results of the Fluid Accumulation Status Trial (FAST) [Abstract]. J Card Fail 2009;15:813. Cleland JG, Louis AA, Rigby AS, Janssens U, Balk AH. Noninvasive home telemonitoring for patients with heart failure at high risk of recurrent admission and death: the Trans-European NetworkeHome-Care Management System (TEN-HMS) study. J Am Coll Cardiol 2005;45:1654e64. Yu CM, Wang L, Chau E, Chan RH, Kong SL, Tang MO, et al. Intrathoracic impedance monitoring in patients with heart failure: correlation with fluid status and feasibility of early warning preceding hospitalization. Circulation 2005;112:841e8. Small RS, Wickemeyer W, Germany R, Hoppe B, Andrulli J, Brady PA, et al. Changes in intrathoracic impedance are associated with subsequent risk of hospitalizations for acute decompensated heart failure: clinical utility of implanted device monitoring without a patient alert. J Card Fail 2009;15:475e81. Vollmann D, Nagele H, Schauerte P, Wiegand U, Butter C, Zanotto G, et al. Clinical utility of intrathoracic impedance monitoring to alert patients with an implanted device of deteriorating chronic heart failure. Eur Heart J 2007;28:1835e40. Anand IS, Greenberg BH, Fogoros RN, Libbus I, Katra RP. Design of the Multi-sensor Monitoring in Congestive Heart Failure (MUSIC) study: prospective trial to assess the utility of continuous wireless physiologic monitoring in heart failure. J Card Fail 2011;17:11e6. Shultz EK. Multivariate receiver-operating characteristic curve analysis: prostate cancer screening as an example. Clin Chem 1995;41:1248e55. Klose CD, Klose AD, Netz U, Beuthan J, Hielscher AH. Multiparameter classifications of optical tomographic images. J Biomed Opt 2008; 13:050503. Fonarow GC, Stough WG, Abraham WT, Albert NM, Gheorghiade M, Greenberg BH, et al. Characteristics, treatments, and outcomes of patients with preserved systolic function hospitalized for heart failure: a report from the OPTIMIZE-HF registry. J Am Coll Cardiol 2007; 50:768e77. Solomon SD, Anavekar N, Skali H, McMurray JJ, Swedberg K, Yusuf S, et al. Influence of ejection fraction on cardiovascular outcomes in a broad spectrum of heart failure patients. Circulation 2005;112:3738e44. Chaudhry SI, Wang Y, Concato J, Gill TM, Krumholz HM. Patterns of weight change preceding hospitalization for heart failure. Circulation 2007;116:1549e54. Sweitzer NK, Lopatin M, Yancy CW, Mills RM, Stevenson LW. Comparison of clinical features and outcomes of patients hospitalized with

Multisensor Monitoring in CHF heart failure and normal ejection fraction ($55%) versus those with mildly reduced (40% to 55%) and moderately to severely reduced (!40%) fractions. Am J Cardiol 2008;101:1151e6. 19. Yancy CW, Lopatin M, Stevenson LW, de Marco T, Fonarow GC. Clinical presentation, management, and in-hospital outcomes of patients admitted with acute decompensated heart failure with preserved systolic function: a report from the Acute Decompensated Heart



Anand et al

295

Failure National Registry (ADHERE) database. J Am Coll Cardiol 2006;47:76e84. 20. Fonarow GC, Heywood JT, Heidenreich PA, Lopatin M, Yancy CW. Temporal trends in clinical characteristics, treatments, and outcomes for heart failure hospitalizations, 2002 to 2004: findings from Acute Decompensated Heart Failure National Registry (ADHERE). Am Heart J 2007;153:1021e8.