Individualized Weight Monitoring Using the Heartphone
Patients with HF deemed high risk for recurrent clinical events attending the HF Unit at St Vincent's University Healthcare Group were approached for inclusion in this study. The protocol had been reviewed by the local Ethics Committee and adhered to standards laid down by the principles of the Declaration of Helsinki. All patients gave informed consent for inclusion in this study and were under the care of clinical and medical staff in the HF Unit, St Vincent's University Hospital, Dublin, Ireland.
Criteria to define high risk of recurrent clinical events included at least one of the following: recent admission for acute decompensated HF; recent outpatient deterioration requiring i.v. loop diuretic; proven non-compliance with medical therapy defined by documented non-adherence using the validated Moriskys Medication Adherence Scale and/or non-persistence with disease-modifying therapy for HF; or a euvolaemic brain natriuretic peptide (BNP) value >300 pg/mL. Consenting patients underwent remote weight monitoring using the HeartPhone system, and the follow-up period for the purpose of the present data analysis was confined to the period between the dates of the first and last transmitted weights. The HeartPhone system involves transmission of weight from a Bluetooth-enabled weighing scales via software on a mobile phone to a remote web server. Patients and carers were given education emphasizing self-care, with early reporting of symptoms and unexpected weight gain of >2 kg over a period of 2–3 days in accordance with the European guidelines.
Clinical deterioration of HF was defined as evidence of HF decompensation diagnosed at an unscheduled visit to the outpatient HF clinic requiring augmentation of oral or i.v. diuretic treatment, or requiring hospital admission as described in European guidelines. In all cases, diagnosis of clinical deterioration was established by a consultant cardiologist or specialist physician experienced in HF care, based on early reporting of change in symptoms, patient self-report of body weight, physical examination, BNP, and the optional use of chest X-ray in uncertain cases. For the purpose of evaluation of weight changes relative to clinical deterioration of HF, the beginning of deterioration was the first day of diagnosis of the deterioration at the clinic. Patients with confirmed clinical deterioration were treated with diuretic modification (oral or i.v.), and clinical response was assessed with a follow-up visit to the clinic with the option for further alteration in therapy or hospitalization if clinically indicated. All clinical staff and patients were blinded to the details of the HeartPhone algorithm. In the case of 40 consecutive patients, the remotely monitored weight data were not available to the clinician, and in the remainder (n = 47) remotely monitored weight data were available although not necessarily consulted. In the case of one hospitalization for HF, there was a failure to transmit weight for >2 weeks in advance of the hospitalization, and this event was discarded for the present analysis.
All remote daily weights transmitted were screened and checked for compatibility with the patients' weight, and weights above or below a specific threshold of the previous value or missing values were discarded and replaced with a calculated value. This occasionally occurred, for example, if another person other than the monitored patient used the scales. If there was more than one weight transmission per day, only the first valid weight was included. Failure to transmit weight on >1 day prompted an outbound phone call from a HeartPhone technician who addressed the problem (e.g. weight not taken, data not yet transmitted). Adherence with weight monitoring was calculated retrospectively as the percentage of days for which valid data was transmitted divided by the number of monitored days.
We evaluated the changes in population weights from baseline and over a 3-month follow-up period. Population weight was also evaluated in patients who experienced a clinical deterioration of HF over a 2-week period before and after the date of clinic diagnosis of deterioration. In those patients who did not deteriorate, a randomly selected 2-week period was used for weight change comparison. Underlying weight variability was evaluated over the study period using 7-day moving averages as follows: we measured the mean of the 7-day moving average for aggregated population weights and the mean amplitude of 7-day moving averages for individual patient weights. This analysis allowed comparison of the stability of underlying population and individual weights over time.
The HeartPhone weight monitoring algorithm is based on the application of moving averages to daily weight data and the generation of alerts from deviations above the norm for the individual patient, and has been described in more detail in a published patent application PCT/EP2011/052788. The algorithm is described therein as having a number of distinct iterations. In all cases, the threshold used for alerts is a mathematical function of the patients' weight and is automatically generated. One abnormal measurement is sufficient for an alert, although alerts are frequently consecutive and in such instances are considered as a continuous alert. Changes in the underlying weight are tracked using moving averages including when patients are deemed clinically stable. Major deviations are distinguished from minor deviations by the extent of and persistence of deviation from the threshold. Further details of the HeartPhone weight monitoring algorithm can be obtained by contacting the sponsor of the study, the National Digital Research Centre (info@ndrc.ie).
Apart from the data processing outlined above, the algorithm individualizes weight monitoring in three ways: by automatically tracking underlying changes in a patient's dry weight; by adjusting for the individual patient's natural weight variability; and by applying a threshold that is in proportion to the patient's weight rather than applying a single absolute threshold as advocated by guidelines. It determines deviations from the moving averages and distinguishes between different grades of deviations ('minor' and 'major') by evaluating both the magnitude and the duration of deviations. The algorithm advocates patient contact with follow-up for appropriate deviations only, and close monitoring for minor deviations. Finally, the sensitivity and specificity of the algorithm can be adjusted by adjustment of the moving average monitoring period and threshold function.
The sensitivity and specificity of two guideline methods (2 kg over 2–3 days and 1.36 kg over 1 day) and the HeartPhone algorithm for determination of clinical deterioration using weight were calculated and compared using a modification of the methodology of Zhang and colleagues. In the present methodology, the remote monitoring period of each patient was divided into weekly periods, rather than 2 weekly periods as used previously, meaning that when a proven outpatient or inpatient clinical deterioration of HF occurred, the week before diagnosis of deterioration was evaluated for guideline and HeartPhone weight alerts. On average, this approach would tend to reduce the sensitivity and positive predictive value of the present analyses for clinical deterioration compared with previous work. However, we believe it to be more clinically relevant in our data set based on the observation in Figure 2 that weight changes in our population were significantly elevated on average during the week prior to the clinical deterioration of HF. In addition, we calculated the number of weeks during which there was a guideline or HeartPhone alert generated and yet no confirmed clinical deterioration occurred. In the evaluation of HeartPhone, we used two iterations based on different moving average monitoring periods and thresholds. These generated HeartPhone algorithm A and B which were optimized for specificity and sensitivity, respectively. Sensitivity, specificity, and positive and negative predictive values of all approaches were calculated and compared. Finally, we evaluated the impact of adjusting the duration of the moving average window period and patient threshold as well the time dependence of alert generation on false-positive rates. In the latter, we compared the alert frequency in the early and later post-discharge periods.
(Enlarge Image)
Figure 2.
(A) Population weight mean (± SEM) over a 28-day period pre- and post-clinic diagnosis of symptomatic deterioration of heart failure (HF) (day 0) in 19 patients with 28 deterioration events: the increase in weight over baseline (day –14) was significant (all P ≤ 0.01) from days –4 to 0 inclusive. Significant increases over day –14 using a paired sample t-test in the 2-week interval up to day 0 were seen consecutively on days –4 to 0 (day –3 vs. day –14, P = 0.023; all others vs. day –14, P < 0.01). The average weight from the second day post-diagnosis showed significant reductions (all P ≤ 0.01) from the population weight on the day of diagnosis (day 0). (B) Population weight mean (± SEM) over a randomly selected 28-day period in 19 patients with no clinical deterioration of HF. These data are presented for comparison purposes with (A) and there were no significant changes in weight from baseline (day –14) over the follow-up period.
For continuous variables, summary statistics are presented as the mean and standard deviations or median and 25th to 75th percentiles (interquartile ranges) as appropriate unless otherwise stated. Categorical variables are presented as frequencies and percentages (in parentheses). Comparisons were made using independent t-tests, paired sample t-tests, Mann–Whitney tests, or χ tests where appropriate. The sensitivity, specificity, positive predictive value, and negative predictive value of guideline and HeartPhone weight monitoring for clinical deterioration of HF are calculated using standard methodology.
The authors had full access to the data and take responsibility for their integrity. All authors read and agreed to the manuscript as written. All statistics were carried out using SPSS version 11 statistical software.
Methods
Patient Population
Patients with HF deemed high risk for recurrent clinical events attending the HF Unit at St Vincent's University Healthcare Group were approached for inclusion in this study. The protocol had been reviewed by the local Ethics Committee and adhered to standards laid down by the principles of the Declaration of Helsinki. All patients gave informed consent for inclusion in this study and were under the care of clinical and medical staff in the HF Unit, St Vincent's University Hospital, Dublin, Ireland.
Criteria to define high risk of recurrent clinical events included at least one of the following: recent admission for acute decompensated HF; recent outpatient deterioration requiring i.v. loop diuretic; proven non-compliance with medical therapy defined by documented non-adherence using the validated Moriskys Medication Adherence Scale and/or non-persistence with disease-modifying therapy for HF; or a euvolaemic brain natriuretic peptide (BNP) value >300 pg/mL. Consenting patients underwent remote weight monitoring using the HeartPhone system, and the follow-up period for the purpose of the present data analysis was confined to the period between the dates of the first and last transmitted weights. The HeartPhone system involves transmission of weight from a Bluetooth-enabled weighing scales via software on a mobile phone to a remote web server. Patients and carers were given education emphasizing self-care, with early reporting of symptoms and unexpected weight gain of >2 kg over a period of 2–3 days in accordance with the European guidelines.
Clinical Deterioration of Heart Failure
Clinical deterioration of HF was defined as evidence of HF decompensation diagnosed at an unscheduled visit to the outpatient HF clinic requiring augmentation of oral or i.v. diuretic treatment, or requiring hospital admission as described in European guidelines. In all cases, diagnosis of clinical deterioration was established by a consultant cardiologist or specialist physician experienced in HF care, based on early reporting of change in symptoms, patient self-report of body weight, physical examination, BNP, and the optional use of chest X-ray in uncertain cases. For the purpose of evaluation of weight changes relative to clinical deterioration of HF, the beginning of deterioration was the first day of diagnosis of the deterioration at the clinic. Patients with confirmed clinical deterioration were treated with diuretic modification (oral or i.v.), and clinical response was assessed with a follow-up visit to the clinic with the option for further alteration in therapy or hospitalization if clinically indicated. All clinical staff and patients were blinded to the details of the HeartPhone algorithm. In the case of 40 consecutive patients, the remotely monitored weight data were not available to the clinician, and in the remainder (n = 47) remotely monitored weight data were available although not necessarily consulted. In the case of one hospitalization for HF, there was a failure to transmit weight for >2 weeks in advance of the hospitalization, and this event was discarded for the present analysis.
Data Management
All remote daily weights transmitted were screened and checked for compatibility with the patients' weight, and weights above or below a specific threshold of the previous value or missing values were discarded and replaced with a calculated value. This occasionally occurred, for example, if another person other than the monitored patient used the scales. If there was more than one weight transmission per day, only the first valid weight was included. Failure to transmit weight on >1 day prompted an outbound phone call from a HeartPhone technician who addressed the problem (e.g. weight not taken, data not yet transmitted). Adherence with weight monitoring was calculated retrospectively as the percentage of days for which valid data was transmitted divided by the number of monitored days.
Evaluation of Population and Individual Weight Over Follow-up
We evaluated the changes in population weights from baseline and over a 3-month follow-up period. Population weight was also evaluated in patients who experienced a clinical deterioration of HF over a 2-week period before and after the date of clinic diagnosis of deterioration. In those patients who did not deteriorate, a randomly selected 2-week period was used for weight change comparison. Underlying weight variability was evaluated over the study period using 7-day moving averages as follows: we measured the mean of the 7-day moving average for aggregated population weights and the mean amplitude of 7-day moving averages for individual patient weights. This analysis allowed comparison of the stability of underlying population and individual weights over time.
HeartPhone Weight Monitoring Algorithm
The HeartPhone weight monitoring algorithm is based on the application of moving averages to daily weight data and the generation of alerts from deviations above the norm for the individual patient, and has been described in more detail in a published patent application PCT/EP2011/052788. The algorithm is described therein as having a number of distinct iterations. In all cases, the threshold used for alerts is a mathematical function of the patients' weight and is automatically generated. One abnormal measurement is sufficient for an alert, although alerts are frequently consecutive and in such instances are considered as a continuous alert. Changes in the underlying weight are tracked using moving averages including when patients are deemed clinically stable. Major deviations are distinguished from minor deviations by the extent of and persistence of deviation from the threshold. Further details of the HeartPhone weight monitoring algorithm can be obtained by contacting the sponsor of the study, the National Digital Research Centre (info@ndrc.ie).
Apart from the data processing outlined above, the algorithm individualizes weight monitoring in three ways: by automatically tracking underlying changes in a patient's dry weight; by adjusting for the individual patient's natural weight variability; and by applying a threshold that is in proportion to the patient's weight rather than applying a single absolute threshold as advocated by guidelines. It determines deviations from the moving averages and distinguishes between different grades of deviations ('minor' and 'major') by evaluating both the magnitude and the duration of deviations. The algorithm advocates patient contact with follow-up for appropriate deviations only, and close monitoring for minor deviations. Finally, the sensitivity and specificity of the algorithm can be adjusted by adjustment of the moving average monitoring period and threshold function.
Comparison of Guideline Weight Monitoring and the HeartPhone Algorithm
The sensitivity and specificity of two guideline methods (2 kg over 2–3 days and 1.36 kg over 1 day) and the HeartPhone algorithm for determination of clinical deterioration using weight were calculated and compared using a modification of the methodology of Zhang and colleagues. In the present methodology, the remote monitoring period of each patient was divided into weekly periods, rather than 2 weekly periods as used previously, meaning that when a proven outpatient or inpatient clinical deterioration of HF occurred, the week before diagnosis of deterioration was evaluated for guideline and HeartPhone weight alerts. On average, this approach would tend to reduce the sensitivity and positive predictive value of the present analyses for clinical deterioration compared with previous work. However, we believe it to be more clinically relevant in our data set based on the observation in Figure 2 that weight changes in our population were significantly elevated on average during the week prior to the clinical deterioration of HF. In addition, we calculated the number of weeks during which there was a guideline or HeartPhone alert generated and yet no confirmed clinical deterioration occurred. In the evaluation of HeartPhone, we used two iterations based on different moving average monitoring periods and thresholds. These generated HeartPhone algorithm A and B which were optimized for specificity and sensitivity, respectively. Sensitivity, specificity, and positive and negative predictive values of all approaches were calculated and compared. Finally, we evaluated the impact of adjusting the duration of the moving average window period and patient threshold as well the time dependence of alert generation on false-positive rates. In the latter, we compared the alert frequency in the early and later post-discharge periods.
(Enlarge Image)
Figure 2.
(A) Population weight mean (± SEM) over a 28-day period pre- and post-clinic diagnosis of symptomatic deterioration of heart failure (HF) (day 0) in 19 patients with 28 deterioration events: the increase in weight over baseline (day –14) was significant (all P ≤ 0.01) from days –4 to 0 inclusive. Significant increases over day –14 using a paired sample t-test in the 2-week interval up to day 0 were seen consecutively on days –4 to 0 (day –3 vs. day –14, P = 0.023; all others vs. day –14, P < 0.01). The average weight from the second day post-diagnosis showed significant reductions (all P ≤ 0.01) from the population weight on the day of diagnosis (day 0). (B) Population weight mean (± SEM) over a randomly selected 28-day period in 19 patients with no clinical deterioration of HF. These data are presented for comparison purposes with (A) and there were no significant changes in weight from baseline (day –14) over the follow-up period.
Statistical Analyses
For continuous variables, summary statistics are presented as the mean and standard deviations or median and 25th to 75th percentiles (interquartile ranges) as appropriate unless otherwise stated. Categorical variables are presented as frequencies and percentages (in parentheses). Comparisons were made using independent t-tests, paired sample t-tests, Mann–Whitney tests, or χ tests where appropriate. The sensitivity, specificity, positive predictive value, and negative predictive value of guideline and HeartPhone weight monitoring for clinical deterioration of HF are calculated using standard methodology.
The authors had full access to the data and take responsibility for their integrity. All authors read and agreed to the manuscript as written. All statistics were carried out using SPSS version 11 statistical software.
SHARE
