This is an openaccess article distributed under the terms of the Creative Commons Attribution License (http://creativecommons.org/licenses/by/2.0/), which permits unrestricted use, distribution, and reproduction in any medium, provided the original work, first published in JMIR Medical Informatics, is properly cited. The complete bibliographic information, a link to the original publication on http://medinform.jmir.org/, as well as this copyright and license information must be included.
Heart Failure (HF) is a common reason for hospitalization. Admissions might be prevented by early detection of and intervention for decompensation. Conventionally, changes in weight, a possible measure of fluid accumulation, have been used to detect deterioration. Transthoracic impedance may be a more sensitive and accurate measure of fluid accumulation.
In this study, we review previously proposed predictive algorithms using body weight and noninvasive transthoracic bioimpedance (NITTI) to predict HF decompensations.
We monitored 91 patients with chronic HF for an average of 10 months using a weight scale and a wearable bioimpedance vest. Three algorithms were tested using either simple ruleofthumb differences (RoT), moving averages (MACD), or cumulative sums (CUSUM).
Algorithms using NITTI in the 2 weeks preceding decompensation predicted events (
NITTI measurements decrease before decompensations, and combined with trend algorithms, improve the detection of HF decompensation over current guideline rules; however, many alerts are not associated with clinically overt decompensation.
Chronic heart failure (HF) is common [
Worsening heart failure may lead to weight gain as a consequence of fluid retention and edema and, if uncorrected, can lead to hospitalization and ultimately death. The Heart Failure Association of America (HFSA) and the European Society of Cardiology (ESC) guidelines both recommend daily weight monitoring. The ESC recommends that patients experiencing a weight increase of 2 kg or more in 3 days should alert healthcare professionals and increase their diuretic dose [
Worsening hemodynamics with increased vascular resistance, afterload mismatch, congestion, and diastolic dysfunction are thought to precede fluid accumulation [
An increased risk of decompensation has been shown for both weight gain [
The aim of this investigation was to evaluate and compare the predictive value of previously published algorithms using measurements of daily body weight, and noninvasive measures of NITTI from a smarttextile vest, to detect decompensation prior to the onset of severe symptoms leading to hospitalization.
The data for this analysis were collected as part of the MyHeart heart failure management observational study [
Of 148 patients recruited from October 2008 to July 2010, 108 had the system installed and data recorded; 3 did not fit the criteria, 3 were unavailable at installation, 1 died before installation, and 33 withdrew before system installation. Of the remaining 108 users, 17 used the system on less than 30 occasions, leaving 91 patients as the focus of this exploratory analysis. Their mean (SD) age was 63 (12) years and 64 were men. Mean weight was 84 (19) kg, mean BMI was 29 (6) kg/m^{2}, and mean left ventricular ejection fraction (LVEF) was 31 (12) %. Most patients had mild (NYHA class II: 60%) or moderate (NYHA class III: 36%) symptoms. Etiology was ischemic in 47%, idiopathic dilated cardiomyopathy in 31%, valvular disease in 5%, and other in 9%. Comorbidities included hypertension (68%), diabetes (37%), atrial fibrillation (36%), renal dysfunction (28%) and COPD (13%). Treatment included angiotensin converting enzyme (ACE) or angiotensin receptor blockers (ARB) (87%), betablockers (88%), MRA (53%), diuretics (84%), digoxin (21%), and implantable cardioverterdefibrillator/cardiac resynchronization therapy (ICD/CRT) (23%/14%). The average monitoring time was 10 months, during which 19 patients were hospitalized one or more times due to decompensated HF, with a total of 24 decompensated HF hospitalizations. The adverse events were adjudicated by an advisory committee.
Patients were instructed on how to perform measurements of body weight and NITTI. Measurements were carried out in the morning before eating breakfast. Body weight was collected using a weight scale (Philips Medical Systems, Andover, Massachusetts, USA), which automatically logged the measurements (accuracy ± 0.1 kg). TTI was measured using a wearable bioimpedance vest [
The bioimpedance vest shown by a model subject correctly applying it across the chest. Textile electrodes on each side of the flexible measurement panel inject currents at different frequencies and register the resulting voltage to calculate the impedance parameter relating to extracellular fluid volume.
The weight and NITTI data were applied to published algorithms (detailed description in
The predictive power of the algorithms was assessed by exploring their ability to alarm within a prespecified period before a hospitalization due to worsening heart failure. Changes in NITTI are thought to precede changes in weight prior to hospitalization [
Three types of alert algorithms are compared in this study: ruleofthumb (RoT) [
1.
2.
3.
4.
Generated example data with the underlying trend in NITTI are shown in the top graph. The resulting output of the three algorithms, normalized to the last measure to show the qualitative difference between the algorithms, is shown in the bottom graph.
Each of the algorithms considered in this study (RoT, MACD, CUSUM) have modifiable parameters that will alter their behavior and ultimately their predictive performance. We tested the performance of each algorithm for a range of possible parameter values. For the RoT algorithms, the number of days (
Segmentation of the data into 2week periods results in substantially more periods without an HFrelated hospitalization compared to those with one. To avoid producing algorithms that raise a large number of false positive alarms, previous studies have focused only on alarms with high specificity [
Parameter optimization can lead to models that overfit the data, which then would not generalize well to other data sets. To minimize these effects, we implemented a stratified leavepatientout crossvalidation (CV) method for the parameters in the RoT, MACD, and CUSUM algorithms. This procedure randomly splits the data into 8 groups, while maintaining the number of patients and decompensation events in each group. The parameters were then optimized for the data with one group left out. The data from the leftout group were then used to evaluate the performance of the optimized parameters. This was repeated until all groups had been left out once. The leftout groups were then recombined to provide an unbiased ROC curve. The optimal threshold for the output index was chosen to be the Youden point with specificity larger than 90%.
Comparisons between the recorded measurements and the output index for the different algorithms in the 2 weeks preceding hospitalization and all other periods were tested with a mixedeffect model using patient specific intercepts as random effects. An arbitrary significance of 0.05 was assumed throughout. Missing data due to adherence issues were removed from the analysis by excluding periods in which less than 3 [
Among the 91 patients for whom data were included in the analysis, 24 heart failurerelated hospitalizations occurred in 19 patients. Of the 24 hospitalizations, 9 had less than 3 weekly weight recordings and 12 had less than 3 weekly impedance recordings preceding the hospitalization, and were excluded from the analysis. The minimum window for the CUSUM algorithm excluded an additional 2 for its analysis.
The predictive performance of guidelinebased rules and published algorithms using weight are presented in
The crossvalidation analyses of the developed models based on published algorithms are presented in
The MACD algorithm improved performance for both weight and impedance. The CUSUM algorithm improved performance for NITTI. The performance of trend algorithms was superior to previously published algorithms (
Performance of different weight algorithms in anticipating an upcoming decompensation.
Source  Weight algorithm  Sensitivity 
Specificity 
PPV^{a}

NPV^{b}

Guideline issuing bodies  >2 lbs^{c} in 1 day [ 
67  56  1.4  99.5 

>2 kg in 3 days [ 
13  87  0.9  99.1 

>4 lbs^{c} in 1 week [ 
27  87  1.8  99.2 
Existing literature  Random chance  50  50  0.9  99.1 

>2 lbs in 1 day or >3 lbs in 3 days [ 
73  50  1.3  99.5 

>2 lbs in 1 day or >5 lbs in 3 days [ 
67  56  1.4  99.4 

>3 lbs in 1 day or >5 lbs in 3 days [ 
13  82  0.7  99.1 

>3 lbs in 1 day or >7 lbs in 3 days [ 
7  83  0.4  99.0 

>4 lbs in 1 day or >7 lbs in 3 days [ 
7  93  0.9  99.1 

>4 lbs in 1 day or >9 lbs in 3 days [ 
7  93  0.9  99.1 

>5 lbs in 1 day or >9 lbs in 3 days [ 
0  100  —  99.1 

>2 lbs in 1 week [ 
80  45  1.3  99.6 

>5 lbs in 1 week [ 
20  94  2.7  99.2 

>4 lbs in a 5 to 80 days MACD^{d} [ 
20  97  6.3  99.3 
^{a}PPV: positive predictive value
^{b}NPV: negative predictive value
^{c}To convert to kilograms multiply by 0.45
^{d}MACD: moving average convergence divergence
ROC curves from the crossvalidated evaluation for the three considered algorithms in the specificity range from 0.9 to 1. A shows the rule of thumb algorithm, B the MACD algorithm, and C the CUSUM algorithm. Performance using NITTI measures is shown with the dashed green line, weight is shown with the blue line, and random chance is portrayed by the red dotted line.
The output of the 2 best performing algorithms for weight and impedance with optimal parameters (maximum Youden index with specificity >90%) is shown in
Crossvalidated performance measures of the algorithms at the maximum Youden index within a specificity of 90100%.
Optimal algorithms^{a}  Sensitivity 
Specificity 
PPV^{b}

NPV^{c}









RoT^{d}: >2.7 kg in 17 days  20  90  1.95  99.2 

MACD^{e}: >0.62 kg (N_{s}=9, N_{l}= 20 days)  33  91  3.2  99.3 

CUSUM^{f}: >8.7 with 10day average, c=0.75  13  91  1.4  99.1 







RoT: <0.27 (log ohm) in 21 days  33  92  4.2  99.2 

MACD: <0.059 (log ohm) (N_{s}=9, N_{l}= 35 days)  50  92  5.9  99.5 

CUSUM: <7.8 with 20day average, c=0.75  60  96  10.9  99.6 
^{a}The optimal parameters and thresholds were estimated from the full data (for stability and variance of crossvalidated parameters and thresholds, see
^{b}PPV: positive predictive value
^{c}NPV: negative predictive value
^{d}RoT: rule of thumb
^{e}MACD: moving average convergence divergence
^{f}CUSUM: cumulative sums
^{g}NITTI: noninvasive transthoracic bioimpedance
Three weeks of telemonitoring data from two patients with high compliance before an upcoming decompensation. Circles correspond to NITTI measurements and the NITTICUSUM algorithm and crosses correspond to weight measurements and the weightMACD algorithm. Optimal thresholds are shown as dashdotted lines in green for NITTI and dotted blue lines for weight.
The use of a crossvalidation procedure to minimize biased performance measures generated several plausible parameters for the tested algorithms; these are presented in
Mean values for weight, impedance, and the respective output indices of the optimal algorithms during periods preceding a hospitalization compared to the other periods are shown in
Mean, standard deviation, and individual values for the estimated optimal parameters in each of the 8 folds created using the described stratified crossvalidation procedure.
Measure 

Body weight  Transthoracic impedance  
CV ^{a} step 

1  2  3  4  5  6  7  8  1  2  3  4  5  6  7  8 




Threshold  3.5 (0.08)  0.31 (0.035)  

3.5  3.56  3.4  3.56  3.45  3.6  3.4  3.4  0.3  0.31  0.3  0.3  0.3  0.3  0.3  0.4  

Days  14.4 (3.7)  20.5 (1.41)  

11  17  11  17  17  20  11  11  21  17  21  21  21  21  21  21  




Threshold  0.8 (0.38)  0.10 (0.014)  

1.59  0.62  0.31  0.62  0.62  0.97  0.62  0.95  0.12  0.1  0.1  0.1  0.1  0.09  0.09  0.13  

Shortterm avg. window  8.6 (1.19)  8.1 (0.99)  

8  9  8  9  9  10  9  9  9  8  8  8  8  9  9  6  

Longterm avg. window  25.6 (10.84)  36.3 (3.54)  

50  20  15  20  25  30  20  25  45  35  35  35  35  35  35  35  




Threshold  11.0 (7.87)  8.13 (2.65)  

30  8.7  8.7  8.7  6.9  8.1  8.7  8.1  7.8  10.3  7.8  7.8  11.1  4.40  11.14  4.64  

Days  26.9 (18.3)  18.8 (2.31)  

50  10  10  10  40  40  10  45  20  20  20  20  15  20  15  20  

Depreciation  1.13 (0.40)  0.75 (0.19)  

1.5  0.75  0.75  0.75  1.5  1.5  0.75  1.5  0.75  0.75  0.75  0.75  0.50  1  0.50  1 
^{a}CV: Crossvalidation
^{b}RoT: rule of thumb
^{c}MACD: moving average convergence divergence
^{d}CUSUM: cumulative sums
Population mean output index values for RoT, MACD, and CUSUM algorithms using the optimal parameters (see
Measure  Mean (SD) value in 2week period before decompensation  Mean (SD) value in nondecompensation periods  Statistical significance^{d} 
Weight (kg)  83 (10)  84 (19)  .97 
WeightRoT ^{a} (kg)  0.3 (1.2)  0.06 (0.87)  .76 
WeightMACD ^{b} (kg)  0.08 (0.30)  0.02 (0.22)  .24 
WeightCUSUM ^{c} (kg)  1.9 (2.7)  0.8 (1.3)  .58 
TTI (log Ohm)  3.0 (0.3)  3.4 (0.3)  <.001 
TTIRoT (log Ohm)^{a}  0.07 (0.12)  0.00 (0.08)  <.001 
TTIMACD (log Ohm)^{a}  0.032 (0.044)  0.003 (0.028)  <.001 
TTICUSUM (log Ohm)^{a}  6.4 (9.4)  0.7 (2.0)  <.001 
^{a}RoT: rule of thumb
^{b}MACD: moving average convergence divergence
^{c}CUSUM: cumulative sums
^{d}Estimated with a mixedeffect model with patient specific random effects. For the algorithms the crossvalidation output was used.
The main finding of the present study is that change in NITTI is a stronger predictor of an impending decompensation compared to changes in weight (crossvalidation estimate was 60% for NITTICUSUM vs 33% for WeightMACD) and that both measurements benefit from trend detection algorithms. Mean values of NITTI in the 2week period preceding a decompensation event were lower than in nondecompensation periods (
Fluid overload is one of the leading causes for HF hospitalization and body weight increase has been linked to an increased risk of hospitalization [
The increase in thoracic fluid due to congestion should decrease impedance measurements. Several studies have reported positive results from algorithms using impedance to detect decompensations [
Comparisons between predicted performances of weight and impedance measurements in
The difficulty in assessing prediction algorithms is known [
Therefore, the positive predictive value of 10.9% should be seen in the context of 2week windows having both high specificity and sensitivity and compared to the relatively low predictive value of current weight algorithms.
Low levels of positive predictive value have also been observed in many other studies evaluating prediction algorithms from daily measurements [
Although clinicians were blinded to the observational data, they could have intervened based on increased weight data for worsening patients. If such interventions did not result in a hospitalization, they were not recorded in this study and might have negatively affected the results. In the SENSEHF trial [
Incorrectly using the measurement equipment could have caused erroneous values with the net effect of lowered performances. The surface on which the scales sit, their accuracy, clothing, and use by other family members can all cause problems with measurement. Bioimpedance weight scales (a different technology from NITTI) require patients to remove their socks and shoes and hence may improve the consistency of measurement. Giving patients feedback and asking them to recheck their weight if it falls out of the expected range are all likely to improve the data quality on which the algorithms are based. The limited amount of data available for this study makes generalizations difficult. Application of crossvalidation procedures were employed to minimize this effect; however, the calculated percentage values were ultimately derived from a small set of subjects and should therefore be seen as qualitative indicators of performance.
Daily measurements of transthoracic impedance using a vest with textile electrodes is a feasible way to monitor HF and provides a more accurate indication of upcoming decompensations when compared to weight for all 3 algorithms tested (RoT, MACD, and CUSUM). Trend detection algorithms outperformed RoT measures suggesting that tracking the progression is more important than direct measures of change, which currently are suggested by guidelines.
However, the low positive predictive value of all the algorithms tested did not allow accurate prediction of impending HF hospitalizations. Implementation of trend detection algorithms might better serve as indications of worsening, which, when integrated with other clinical measures, could be useful for treatment management. The promising results from this investigation warrant further trials with noninvasive TTI as a technology for the management of HF, perhaps connected to actionable alerts. These alerts would promote a strategy of “health maintenance” to keep the patient as close to their ideal state as possible on a daily basis, which could be combined with a strategy of “crisis detection and management” if the first strategy failed.
Detailed description of algorithms to detect decompensated HF.
angiotensin converting enzyme
angiotensin receptor blockers
coronary artery bypass grafting
cumulative sums
crossvalidation
European Society of Cardiology
Heart Failure
Heart Failure Association of America
implantable cardioverterdefibrillator/cardiac resynchronization therapy
intrathoracic impedance
left ventricular ejection fraction
moving average convergence divergence
New York Heart Association
noninvasive transthoracic bioimpedance
negative predictive value
pulmonary arterial pressure
positive predictive value
receiver operator curve
ruleofthumb
This work was supported by the EU Marie Curie Network iCareNet under grant number 264738. Data were provided by the MyHeart project, which was partially financed by the EU FP6 program under grant number 507816.
ICG is a PhD student employed at Philips Research. AGB, HR, and JH are employed by Philips Research. JGFC and KGM have received departmental research support from Philips.