1Universitätsklinikum Würzburg Deutsches Zentrum für Herzinsuffizienz Würzburg, Deutschland; 2ZANA Technologies GmbH Kalsruhe, Deutschland; 3University Hospital Würzburg Department of Otorhinolaryngology, Plastic, Aesthetic and Reconstructive Head and Neck Surgery Würzburg, Deutschland; 4Universitätsklinikum Würzburg Medizinische Klinik I, Kardiologie Würzburg, Deutschland; 5Cosinuss GmbH Munich, Deutschland; 6Universitätsklinikum Würzburg Medizinische Klinik und Poliklinik I Würzburg, Deutschland
Introduction Heart failure (HF) is a frequent yet complex condition with a diverse etiology and large clinical heterogeneity. We explored the application of digital vocal biomarkers with unsupervised machine learning in order to identify novel HF subgroups that may aid personalizing monitoring and treatment.
Methods The AHF-Voice study is a BMBF-funded ongoing monocentric prospective cohort study conducted at the University Hospital Würzburg. Inclusion criteria are hospitalization for acute HF, age ≥18 years, life-expectancy ≥6 months. Exclusion criteria are high output HF, cardiogenic shock, listing for high-urgency heart transplantation or a history of vocal fold disease or phonosurgery. Supervised by study staff, patients collect daily voice recordings in a dedicated smartphone app using three different voice tasks: spontaneous speech, sustained vowel, text reading. Patients are comprehensively phenotyped during hospitalization. We set up a vocal biomarker pipeline for audio processing and extraction of phonation and spectral-based voice features. Data were analyzed using principal component analysis (PCA) with an unsupervised K-means clustering approach. The number of clusters was determined using the Silhouette score. The AHF-Voice aims to recruit 123 patients and follow them for 6 months. We here report on the in-hospital period of the first 50 AHF-Voice patients included between April and August 2023.
Results Out of 50 patients, 4 withdrew their consent and another 4 patients were excluded due to limited recording quality. Hence, 2753 voice recordings of 42 patients were considered for analysis: mean age was 74±11 and 64% were men. Unsupervised clustering solely based on voice features identified three clusters with distinct phenotypes. Cluster 1 had the longest duration of HF, highest levels of natriuretic peptides and lowest left ventricular ejection fraction (LVEF). Cluster 2 had mid-range LVEF and highest level of potassium. Cluster 3 had the highest LVEF and highest proportion of women.
Conclusion Machine learning-based cluster analysis based on voice features is able to identify distinct groups of HF patients. The clinical utility needs to be explored in the future.
|
1 |
2 |
3 |
P value | |
|
Age (yrs) |
74±8 |
74±13 |
74±11 |
0.967 |
|
Male sex |
7 (78) |
7 (70) |
13 (57) |
0.490 |
|
HF Charateristics | ||||
|
History of HF |
|
|
|
0.009 |
|
de novo |
1 (11) |
7 (70) |
10 (44) |
|
|
<1 years |
- |
- |
3 (13) | |
|
1-5 years |
- |
3 (30) |
2 (9) | |
|
>5 years |
7 (78) |
- |
7 (30) | |
|
unknown |
1 (11) |
- |
1 (4) | |
|
NYHA class III/IV |
8 (89) |
9 (90) |
20 (87) |
0.966 |
|
Comorbidities and risk factors | ||||
|
History of MI |
4 (44) |
3 (30) |
7 (30) |
0.733 |
|
Current smoking |
- |
1 (10) |
4 (17) |
0.393 |
|
Diabetes mellitus |
5 (56) |
5 (50) |
8 (35) |
0.502 |
|
pAVK |
1 (11) |
4 (40) |
4 (17) |
0.250 |
|
COPD |
3 (33) |
1 (10) |
4 (17) |
0.423 |
|
Revascularization |
3 (33) |
4 (40) |
6 (26) |
0.724 |
|
Valve intervention |
3 (33) |
- |
3 (13) |
0.119 |
|
Device (CRT/ICD) |
4 (44) |
- |
8 (35) |
0.067 |
|
Charlson comorbidity score |
3 [2, 5] |
3 [2, 3] |
2 [1, 4] |
0.685 |
|
Measurements | ||||
|
LVEF (%) |
37±12 |
47±14 |
50±16 |
0.052 |
|
BMI (kg/m2) |
31±6 |
32±5 |
31±7 |
0.911 |
|
Systolic pressure (mmHg) |
124±34 |
142±26 |
133±21 |
0.316 |
|
NT-proBNP (pg/mL) |
7387 [2199, 12069] |
4429 [1441, 13894] |
5190 [3100, 11034] |
0.867 |
|
Sodium (mmol/l) |
138 [137, 140] |
138 [138, 140] |
141 [138, 143] |
0.096 |
|
Potassium (mmol/l) |
4±1 |
5±1 |
4±1 |
0.020 |
|
Hematocrit (%) |
39±5 |
36±8 |
38±7 |
0.634 |
|
eGFR (ml/min/1.73m2) |
43 [40, 46] |
66 [45, 85] |
41 [33, 76] |
0.165 |
|
C-reactive protein (mg/dl) |
1 [0, 3] |
1 [0, 1] |
1 [0, 3] |
0.544 |
Data are n (%), mean±SD or median (quartiles). Group comparisons by Kruskal-Wallis test or X² test