Periodic Reporting for period 1 - MATER (Monitoring mentAl healTh in brEast canceR)
Berichtszeitraum: 2023-09-01 bis 2025-08-31
In 2020, in Europe (EU-27), 355,457 women were newly diagnosed with breast cancer and 91,826 women died from it. With an estimated lifetime risk of developing breast cancer of 1 in 7, breast cancer is a public health issue with a heavy impact on patients’ mental health due to the diagnosis implications, treatment or quality-of-life (QoL) interference of the pathology. Patients thus have the double burden of both cancer treatments and/or after-effects, and mental health disorders. In order to enhance patients’ and survivors’ QoL – which is one of the listed priorities of the Europe’s Beating Cancer Plan, to measure adherence to their treatment, and to prevent relapse or the emergence of co-morbidity, these patients require a personalized and regular medical follow-up.
In line with the EU4Health Programme and the e-health European policy, an innovative way to collect these symptoms under a patients’ usual living conditions is the Ecological Momentary Assessment (EMA). EMA consists in measuring parameters related to these symptoms on a very regular basis and under ecological conditions. This collection can be done through questionnaires, applied e.g. via smartphone applications or animated virtual assistants, but also using voice. Indeed, voice is related to the physiological state of the speaker; can be implemented in passive data collection (i.e. data collection that does not require the active participation of the subject – e.g. during a phone call); does not require large computational resources; and is robust to noise, allowing its implementation in diverse environments. Moreover, voice is easily captured through smartphones, and it is estimated that 80% of the world population has access to it : the deployment of this measurement tool is already effective in the general population.
In addition, in order to model the interactions between the different symptoms related to mental health impairment associated with cancer, recent works have used symptom networks. Introduced in 2013 by Borsboom , symptom networks allow to very efficiently visualize the relations between symptoms (e.g. partial correlation) using graphs and to identify the symptoms responsible for the maintenance of a healthy or pathological state (central symptom). Putting together symptoms from different syndromes or disorders into the same graph allows the identification of bridge symptoms , implied in the development of comorbidities, which are of particular interest for EMA. An example of such a network is proposed in Figure 1A. The objective of the MATER project is to take benefits of both the use of voice as a means of measuring symptoms and symptom networks as tools for modeling their interactions, in order to automatically estimate mental health issues in women with breast cancer.
In line with the EU4Health Programme and the e-health European policy, an innovative way to collect these symptoms under a patients’ usual living conditions is the Ecological Momentary Assessment (EMA). EMA consists in measuring parameters related to these symptoms on a very regular basis and under ecological conditions. This collection can be done through questionnaires, applied e.g. via smartphone applications or animated virtual assistants, but also using voice. Indeed, voice is related to the physiological state of the speaker; can be implemented in passive data collection (i.e. data collection that does not require the active participation of the subject – e.g. during a phone call); does not require large computational resources; and is robust to noise, allowing its implementation in diverse environments. Moreover, voice is easily captured through smartphones, and it is estimated that 80% of the world population has access to it : the deployment of this measurement tool is already effective in the general population.
In addition, in order to model the interactions between the different symptoms related to mental health impairment associated with cancer, recent works have used symptom networks. Introduced in 2013 by Borsboom , symptom networks allow to very efficiently visualize the relations between symptoms (e.g. partial correlation) using graphs and to identify the symptoms responsible for the maintenance of a healthy or pathological state (central symptom). Putting together symptoms from different syndromes or disorders into the same graph allows the identification of bridge symptoms , implied in the development of comorbidities, which are of particular interest for EMA. An example of such a network is proposed in Figure 1A. The objective of the MATER project is to take benefits of both the use of voice as a means of measuring symptoms and symptom networks as tools for modeling their interactions, in order to automatically estimate mental health issues in women with breast cancer.
Three main outcomes were achieved during the course of this project, based on the Colive Voice database, which contains both clinical data and voice recordings from targeted participants, i.e. women breast cancer survivors.
1) First, we conducted a network analysis of symptoms using five questionnaires related to fatigue (FSS), mental health (PHQ-9), respiratory health (VQ-11), well-being (WHO-5), and stress (1 question) in a population of 410 breast cancer survivors. We applied all the most recent network analysis methods: calculating symptom centrality to identify core symptoms and bridge symptoms; clustering on the graphs to identify underlying symptom communities; and syndrome networks (i.e. symptom communities).
2) Furthermore, we developed a pipeline for extracting vocal features from the voice recordings. We automatically extracted three groups of descriptors: reading errors at the word level, as well as at the phoneme and phonemic category levels; articulatory rate features such as pauses (duration, number), word and phoneme durations, and the temporal evolution of energy and spectral purity throughout the sentences (slope); and finally, metrics related to vowel articulation, particularly the vowel triangle, which we automatically estimated using a phoneme recognition and segmentation system.
Since the recordings were made in ecological conditions (i.e. in diverse acoustic environments) and with various devices, we had to account for differences in the audio sample quality. To address this, after calculating the audio descriptors, we examined the outliers in the distributions of each feature group to assess the impact of sample defects (cracks, noise, etc.) on the features.
3) Finally, we proceeded with identifying vocal biomarkers of the symptoms collected in the database, accounting for the sensitivity of these descriptors to the various symptoms while considering a number of covariates such as age, education level, and characteristics related to treatments or the participants' medical history.
1) First, we conducted a network analysis of symptoms using five questionnaires related to fatigue (FSS), mental health (PHQ-9), respiratory health (VQ-11), well-being (WHO-5), and stress (1 question) in a population of 410 breast cancer survivors. We applied all the most recent network analysis methods: calculating symptom centrality to identify core symptoms and bridge symptoms; clustering on the graphs to identify underlying symptom communities; and syndrome networks (i.e. symptom communities).
2) Furthermore, we developed a pipeline for extracting vocal features from the voice recordings. We automatically extracted three groups of descriptors: reading errors at the word level, as well as at the phoneme and phonemic category levels; articulatory rate features such as pauses (duration, number), word and phoneme durations, and the temporal evolution of energy and spectral purity throughout the sentences (slope); and finally, metrics related to vowel articulation, particularly the vowel triangle, which we automatically estimated using a phoneme recognition and segmentation system.
Since the recordings were made in ecological conditions (i.e. in diverse acoustic environments) and with various devices, we had to account for differences in the audio sample quality. To address this, after calculating the audio descriptors, we examined the outliers in the distributions of each feature group to assess the impact of sample defects (cracks, noise, etc.) on the features.
3) Finally, we proceeded with identifying vocal biomarkers of the symptoms collected in the database, accounting for the sensitivity of these descriptors to the various symptoms while considering a number of covariates such as age, education level, and characteristics related to treatments or the participants' medical history.
These three main outcomes have led to the generation of the following groups of scientific knowledge:
1) In the network analysis, we identified 8 different clusters corresponding to (1) Depression, (2) Fatigue-related activity limitations, (3) Fatigue-related physical impairment, (4) Fatigue-related participation restrictions, (5) Well-being vitality, (6) Impact on everyday functioning, (7) Cognition & Motor, and (8) Respiratory disability. We identified three bridge symptoms: 'feeling incapable of carrying out projects' (VQ11-6) bridges the 'Fatigue-related participation restrictions' and 'Impact on everyday functioning' syndromes; 'Feeling bad about yourself' (PH9-6) bridges the 'Depressive syndrome' and 'Cognition & Motor' syndromes; and fatigue frequency (PH9-4) bridges the fatigue-related syndromes and depressive syndromes.
Interestingly, the symptom network confirms the distinction we made between fatigue severity (measured in terms of intensity or frequency) and fatigue-related disability (in terms of impairment, activity limitations, and participation restrictions): these different characterizations of fatigue play distinct roles in the symptom network, and particular attention to their differentiation should therefore be emphasized in studies investigating fatigue in cancer. Furthermore, we did not find any link between fatigue-related disability and depression: the connection between fatigue and depression occurs through aspects of fatigue severity (intensity and frequency). Finally, we did not find any significant structural differences between the networks based on the treatments received or the duration of survivorship.
2) In identifying vocal biomarkers of fatigue, we aimed to determine the specific sensitivity of these markers to different symptoms, independent of 8 confounding factors (age, Body Mass Index, education level, history of chemotherapy, surgery, radiotherapy, hormone therapy, and whether the participant was still undergoing treatment at the time of recording). We decoupled our features from these factors and observed that reading errors are significantly affected by having received radiotherapy; articulatory flow characteristics are primarily influenced by age; and phonatory articulation characteristics are mainly affected by BMI, age, and hormone therapy.
We then examined the sensitivity of these descriptors to symptoms after accounting for the influence of confounding factors. Among the vocal biomarkers identified for fatigue, we found that reading errors were preferentially related to physical impairment due to fatigue (dim 3) and respiratory quality of life (dim 8); articulatory flow was linked to these same dimensions, as well as to cognitive and motor aspects (dim 7); finally, phonatory articulation quality metrics appeared sensitive to all the measured dimensions, though they were not highly specific to any of them.
1) In the network analysis, we identified 8 different clusters corresponding to (1) Depression, (2) Fatigue-related activity limitations, (3) Fatigue-related physical impairment, (4) Fatigue-related participation restrictions, (5) Well-being vitality, (6) Impact on everyday functioning, (7) Cognition & Motor, and (8) Respiratory disability. We identified three bridge symptoms: 'feeling incapable of carrying out projects' (VQ11-6) bridges the 'Fatigue-related participation restrictions' and 'Impact on everyday functioning' syndromes; 'Feeling bad about yourself' (PH9-6) bridges the 'Depressive syndrome' and 'Cognition & Motor' syndromes; and fatigue frequency (PH9-4) bridges the fatigue-related syndromes and depressive syndromes.
Interestingly, the symptom network confirms the distinction we made between fatigue severity (measured in terms of intensity or frequency) and fatigue-related disability (in terms of impairment, activity limitations, and participation restrictions): these different characterizations of fatigue play distinct roles in the symptom network, and particular attention to their differentiation should therefore be emphasized in studies investigating fatigue in cancer. Furthermore, we did not find any link between fatigue-related disability and depression: the connection between fatigue and depression occurs through aspects of fatigue severity (intensity and frequency). Finally, we did not find any significant structural differences between the networks based on the treatments received or the duration of survivorship.
2) In identifying vocal biomarkers of fatigue, we aimed to determine the specific sensitivity of these markers to different symptoms, independent of 8 confounding factors (age, Body Mass Index, education level, history of chemotherapy, surgery, radiotherapy, hormone therapy, and whether the participant was still undergoing treatment at the time of recording). We decoupled our features from these factors and observed that reading errors are significantly affected by having received radiotherapy; articulatory flow characteristics are primarily influenced by age; and phonatory articulation characteristics are mainly affected by BMI, age, and hormone therapy.
We then examined the sensitivity of these descriptors to symptoms after accounting for the influence of confounding factors. Among the vocal biomarkers identified for fatigue, we found that reading errors were preferentially related to physical impairment due to fatigue (dim 3) and respiratory quality of life (dim 8); articulatory flow was linked to these same dimensions, as well as to cognitive and motor aspects (dim 7); finally, phonatory articulation quality metrics appeared sensitive to all the measured dimensions, though they were not highly specific to any of them.