Utilising AI Analysis of Sounds To prEdict heaRt failurE decOmpensation

Heart failure is a common condition in which the heart is unable to deliver the desired cardiac output either due to a weakened or stiff heart muscle. It affects more than 2% of the UK population (the incidence is around 200,000 cases per annum) resulting in 5% of all the emergency hospital admissions and it consumes approximately 2% of the annual NHS budget (approximately £2 billion per annum). Therefore, heart failure is not only a major driver for hospitalisation but provides the leading opportunity to reduce preventable admissions.

Acute decompensated heart failure (ADHF) is a medical emergency requiring urgent attention. It usually results in inpatient hospitalisation and is a major driver for associated healthcare costs. ADHF is usually characterised by rapid deterioration of breathlessness at rest or exertion because of pulmonary oedema (pulmonary venous congestion), and fluid retention resulting in swollen legs as well as a myriad of other symptoms including fatigue, lack of appetite, and so on.

The patient normally presents with gradual or sudden onset of typical symptoms (breathlessness, fatigue, and fluid accumulation in the legs). After stabilisation and the initial treatment of ADHF, patients enter a plateau phase where the heart remains stable. However, over time, most patients experience multiple episodes of ADHF which typically become longer and separated by shorter intervals. The congestion is related to underlying increased cardiac pressure usually secondary to volume overload which plays a central role in the pathophysiology, presentation, and prognosis of heart failure. Pulmonary congestion is one of the most important diagnostic and therapeutic targets in heart failure. Detecting pulmonary congestion earlier on due to volume overload is key to preventing impending rehospitalisation and presents an ideal opportunity to optimise heart failure treatment in the community.

Early community detection of ADHF is ultimately the first step in providing effective patient care. Poor recognition of HF due to its multitude of vague/non-specific symptomatology of presentations often leads to delays in diagnosis and treatment. The delay between a patient developing symptoms of HF decompensation and seeking medical attention is often considerable and is influenced by the speed of onset and severity of the symptoms. Therefore, a reliable and easily accessible means of assessing chronic fluid status in ambulatory outpatients is needed to detect early decompensation when appropriate intervention is possible. The sudden development of breathlessness (dyspnoea) from the accumulation of fluid in the lungs (acute pulmonary oedema) usually prompts rapid contact with medical services, whereas the gradual appearance of swollen legs and ankles (peripheral oedema) is more likely to be associated with delays in seeking care. The average delay between symptom onset and hospital admission ranged from 2 hours to 7 days. The symptoms of heart failure often develop gradually and appear non-threatening, potentially explaining some of the observed delays in seeking care.

In recent years, several pilot studies demonstrated a relationship between speech biomarkers and the extent of systemic and/or pulmonary congestion in heart failure patients. For example, in 2017, a study of 10 (8 M, 2F) patients with acute decompensated heart failure undergoing inpatient treatment with intravenous diuretic therapy showed that after treatment, patients displayed a higher proportion of automatically identified creaky voice, increased fundamental frequency, and decreased cepstral peak prominence variation, suggesting that speech biomarkers can be early indicators of HF. The study also showed that the severity of HF-related oedema required to measurably change the voice is small compared to the severity needed to increase body weight, suggesting that speech biomarkers could become a more effective non-invasive tool to monitor HF patients than daily weights. In 2021, another study evaluated the feasibility of remote speech analysis in the evaluation of dynamic fluid overload in heart failure patients undergoing hemodynamic treatment. They performed serial speech/voice measurements in 5 patients undergoing haemodialysis. The analysis was done with an app that does not share its AI algorithm. They demonstrated statistically significant differences in select speech biomarkers at different fluid status levels as the patients progressed through the treatment. Subsequently, in 2022, a comparison of sound recordings for patients admitted with ADHF on the day of admission and the day of discharge with a sample of 40 patients who were admitted with acute decompensated heart failure identified significant differences in all 5 tested speech measures of wet (admission) vs dry (discharge) recordings.

Separately, in 2022, a study evaluated speech and pause alterations in voice recordings of acute (N=68) and stable (N=36) patients and found that the pause ratio was a 14.9% increase in patients of acute HF. They also found a positive correlation with NT-Pro-BNP level. Another study in 2022 examined both Mel-Frequency cepstral coefficient (MFCC) features and glottal speech features, comparing a sample of 25 healthy speakers (7F, 18M) and 20 patients with HF of any aetiology (regardless of LVEF). Following feature selection, they developed predictive models using four different classification methods (SVM, ET, Adaboost, and FFNN). Based on a combination of MFCC and Glottal speech features, they were able to predict ADHF with accuracies ranging from 88-94%, with a true positive rate of 79.47% and true negative rate 82.69%.

By performing an extensive panel of clinical assessments, investigations as well as symptom-based questionnaires in a study involving up to 250 heart failure patients, the investigators aim to build upon recent work and develop a novel AI-based application deployed on a smart device, which can detect an increase in pulmonary congestion from subtle changes in a patient's cough, voice, breathing, and chest sounds. This will provide key information for patients with heart failure and their clinical teams, by correctly detecting progressive fluid accumulation in a patient's lungs prior to the patient developing significant symptoms. Detecting early-phase pulmonary congestion will enable clinicians to target therapy more effectively. It is hoped that this will help minimise and ultimately prevent the need for recurrent emergency hospital admission by alerting the patient to contact their (community) heart failure team and enable earlier outpatient treatment prior to the need to be re-hospitalised entering the acute phase.

Subject to the successful outcome of this research, a prospective interventional clinical trial will then be undertaken, to test the clinical and operational benefits of the AI tool derived from this research on NHS heart failure care, paving the way for the eventual adoption of such solutions in routine clinical practice.