For the first time doctors have shown that measuring changes in 24-hour heart rate can reliably indicate whether or not someone is depressed.
In practical terms, this may give clinicians an objective “early warning” of potential depression, as well as a rapid indication whether or not treatment is working, so opening the way to more rapid and responsive treatment.
Presenting results of this pilot study at the ECNP virtual congress, lead researcher, Dr Carmen Schiweck (Goethe University, Frankfurt) said “Put simply, our pilot study suggests that by just measuring your heart rate for 24 hours, we can tell with 90% accuracy if a person is currently depressed or not”.
Scientists have known that heart rate is linked to depression, but until now they have been unable to understand exactly how one is related to the other.
In part this is because while heart rates can fluctuate quickly, depression both arrives and leaves over a longer period, with most treatments taking months to take effect. This makes it difficult to see whether or not changes in one’s depressive state might be related to heart rate.
“Two innovative elements in this study were the continuous registration of heart rate for several days and nights, and the use of the new antidepressant ketamine, which can lift depression more or less instantly.
This allowed us to see that average resting heart rate may change quite suddenly to reflect the change in mood”, said Carmen Schiweck.
Ketamine has a history as both an anaesthetic and a party drug (a drug of abuse).
However in December last year it was licenced to treat major depression in Europe, after having been introduced in the USA a few months earlier.
Traditional antidepressants can take weeks to show an effect, in contrast ketamine is rapid acting, with results often being seen in minutes.
As Carmen Schiweck said “We knew that something was going on to link heart rate to psychiatric disorders, but we didn’t know what it was, and whether it would have any clinical relevance. In the past researchers had shown that depressed patients had consistently higher heart rates and lower heart rate variability, but because of the time it takes to treat depression it had been difficult to follow up and relate any improvement to heart rate.
But when we realized that ketamine leads to a rapid improvement in mood, we knew that we might be able to use it to understand the link between depression and heart rate”.
Dr Schiweck performed this work in the Mind Body Research group at KU Leuven, Belgium, with Dr Stephan Claes as the principal investigator.
The team worked with a small sample of 16 patients with Major Depressive Disorder, none of who had responded to normal treatment, and 16 healthy controls.
They measured their heartrates for 4 days and 3 nights, and then the volunteers with depression were given either ketamine treatment or a placebo.
“We found that those with depression had both a higher baseline heart rate, and a lower heart rate variation, as we expected. On average we saw that depressed patients had a heart rate which was roughly 10 to 15 beats per minutes higher than in controls.
After treatment, we again measured the heart rates and found that both the rate and the heartrate fluctuation of the previously depressed patients had changed to be closer to those found in the controls”.
The most striking finding was that the scientists were able to use 24-hour heart rate as a “biomarker” for depression. Heart rates were measured using a wearable mini-ECG.
The data was fed to an Artificial Intelligence programme, which was able to classify nearly all controls and patients correctly as being depressed or healthy.
“Normally heart rates are higher during the day and lower during the night. Interestingly, it seems that the drop in heart rate during the night is impaired in depression. This seems to be a way of identifying patients who are at risk to develop depression or to relapse.” said Carmen Schiweck.
The team also found that patients with a higher resting heart rate responded better to the treatment with Ketamine, which may help identify which patients are likely to respond to which treatment.
Carmen Schiweck said “We need to remember that this is a small proof-of-concept study: 6 of our of our 16 initial patients responded to treatment with at least a 30% reduction on the Hamilton Rating scale for depression, so we need to repeat the work with a larger, anti-depressant free sample.
Our next step is to follow up depressed patients and patients who are in remission, to confirm that the changes we see can be used as an early warning system”.
“This is an innovative proof-of-concept study. My own group had previously studied short-term heart rate variability in over a thousand depressed patients and controls, and we did not detect a consistent differentiation, and found antidepressants to have more impact than depression status itself.
However, this study monitored heart rate variability in the ambulatory setting for several days and nights, which gives unique night and day information on the autonomic nervous system. It needs to be examined whether these interesting findings hold in larger, more diverse treatment settings”.
Professor Penninx was not involved in this work, this is an independent comment.
In recent years, small sensors that can be attached to a person’s body for 24 h, known as “wearable devices”, have been widely used (Mazzetta et al., 2018). They can continuously and non-invasively collect a variety of information, including amount of activity, amount of sleep, heart rate, respiratory rate, and physical location (Nieto-Riveiro et al., 2018).
In many medical fields such as cardiology, endocrinology, and metabolic medicine, the use of wearable devices in clinical research and application is growing (Kuehn, 2016; Shelgikar et al., 2016). This trend also holds true in psychiatric research, as data related to activity, sleep, heart rate, etc. have been shown in previous studies to have relevance in determining diagnoses and illness severity (Marzano et al., 2015; Reinertsen and Clifford, 2018).
Because there is a lack of quantifiable biological markers in psychiatry (Beijers et al., 2019), the ability to non-invasively collect data from wearable devices can improve diagnoses and evaluations of illness severity (Dogan et al., 2017; Marzano et al., 2015; Patel et al., 2017).
However, currently there is no sufficient evidence concerning the use of such wearable devices in clinical examinations. Actigraphy, which relies on a device that collects data via an accelerometer, was first used in the psychiatric field as early as the 1970s, and a large number of studies have been conducted since then (Burton et al., 2013; Luik et al., 2015; Martin and Hakim, 2011).
In a meta-analysis focusing on such studies that used actigraphy to evaluate mood disorders (Tazawa et al., 2019), significant differences in the daily activity and sleep-related measurements were found between patients with mood disorders and healthy controls. Moreover, significant differences were found when comparing pre- and post-treatment periods.
Additionally, specific measurement patterns characterizing each mood disorder/status were found.
Thanks to advances in wearable device technology, actigraphy-enabled wearable devices with similar capabilities as those made for research are now available commercially at economical prices.
Moreover, some new modalities that were not measurable with previous devices can now be measured with the new devices, such as heart and respiratory rates, skin temperature, location information, etc. If wearable devices are able to collect multi-modal data at an economical cost, they would be viable tools for evaluating mood disorders in clinical settings, and they could even be used in pre-clinical screenings.
Studies that assessed mood disorders with these new modalities reported the following findings. Moraes et al. (2013) found that in patients with depression (n = 20), there was low amplitude for circadian rhythm and light exposure, and high amplitude for peripheral body temperature compared to healthy controls (n = 10). nullvakmfondeknqsyqiykeegy reported that a weak 24-hour periodicity of body temperature was prominent in melancholic depressed patients (n = 41) compared to healthy controls (n = 25). Licht et al. (2008) reported that those with depression had lower heart rate variability than healthy controls.
In recent years, there has been an increase in studies that analyze mood disorders using various data collected from wearable devices (Rohani et al., 2018; Wang et al., 2018), and in particular, research on bipolar disorder is moving forward (Puiatti et al., 2011; Valenza et al., 2014, 2015).
Valenza et al. used a wearable textile device to analyze heart rate variability, and reported that they were able to predict bipolar disorder patients’ emotional states with an accuracy of over 90%.
Along with the improvement in the sensory abilities of wearable devices, machine learning is becoming increasing popular in the medical field, as clinical data often contain complex cross-sectional and longitudinal patterns. Saad et al. (2019) used heart rate variability during sleep from polysomnograms to distinguish 87 major depressive disorder (MDD) patients and 87 healthy controls utilizing machine learning, then reported a classification accuracy of 79.9%. Valenza et al. (2013) used inter-beat interval time series, heart rate, and respiratory dynamics data to create algorithms to predict mood states.
They collected over 120 h of data from three subjects with bipolar disorder and reported a mood state prediction accuracy of 97%. However, limitations for this study include a small study population of three people, and the fact that no rating scale was used to assign illness and mood state labels. Cho et al. (2019) used activity, sleep, light exposure, and heart rate data from wearable devices and smartphones to predict mood state within the next three days.
They recruited 55 patients with mood disorders, and analyzed their data with machine learning, then reported a prediction accuracy of 64–94%. However, they used their original self-reported mood assessment scale, and predicted the presence of symptoms but not severity.
Given such limitations in previous similar studies, we aimed to investigate the usefulness of wearable devices with sensors for acceleration, heart rate, skin temperature, and ultraviolet (UV) light to identify symptomatic patients as well as measure illness severity through a biostatistical machine learning approach.
In this study, a multimodal wristband-type wearable device was used to estimate the presence and severity of depression. Previous studies have investigated the relationship between depression and each modality measured in this study, but they did not reflect the illnesses in question well enough to be applied as diagnostic tools in clinical practice.
Our research combined multiple datasets with a machine learning approach to create a practical model for estimating both the presence and severity of depressive states.
First, using the seven modalities, we compared each dataset among mood disorder patients and healthy controls. For activity level indicators, we used step count and energy expenditure, which we calculated from accelerometer readings. Both indicators differed significantly between the patient and control study groups.
There are already many studies and meta-analyses that have demonstrated the significant difference between healthy controls and mood disorder patients based on activity levels measured from a three-axis actigraphy device (Teychenne et al., 2008).
On the other hand, step count is an activity marker that is normally measured by a single up/down axis accelerometer called a pedometer, but among studies comparing mood disorder patients and healthy controls, there are few that have measured step count using actigraphy.
McKercher et al. (2009) reported that a pedometer was used to measure the steps taken per day by young adults, and that the prevalence of depression was higher when one’s step count was lower. In our study, we found that healthy controls had significantly greater step counts and energy use during the hours of 11:00 am–6:00 pm, which agrees with the results of previous research.
Regarding sleep, our results showed that sleep time was particularly long among patients during the nighttime hours of 9:00 pm–12:00 am. Insomnia is common in depression (Benca and Peterson, 2008; Riemann and Voderholzer, 2003).
Lack of sleep is very significant and is a major concern when treating depression, but there is also evidence that subjective complaints of insomnia and objective measurements of sleep time do not always align (Argyropoulos et al., 2003).
Our results suggest that objective measurements for sleep showed higher levels of physical calmness during nighttime hours in depressed patients than healthy controls. This may be because patients calm their physical movement earlier in the day compared to healthy controls, or because patients’ social activity levels are lower.
Our results also showed a significant difference between mood disorder patients and healthy controls in the heart rate data collected by the wearable devices’ sensors. In particular, heart rates captured during the sleep hours of 1:00 am–9:00 am show that patients have significantly higher heart rates than healthy controls. Upon investigation, we found numerous studies on depression and heart rate variability (Bassett, 2016; Kemp et al., 2010; Kwon et al., 2019; Stapelberg et al., 2012; Udupa et al., 2007; Wang et al., 2013), but found very little prior research regarding the relationship between heart rate itself and depression.
For example, Kemp et al. (2014) reported that there was no significant difference between depressed patients and healthy controls based on heart rate data taken from 10-minute resting-state ECG tests. Additionally, Carney et al. (2016) reported that when observing depression symptoms in coronary artery disease patients, patients with high heart rates at night had a poorer response to depression treatment.
From a biological standpoint, the reason that a depression patient’s resting heart rate rises may be due to the fact that the automatic nervous system manages one’s resting heart rate.
Based on other studies, it is well known that as stress increases, the hypothalamic-pituitary-adrenal (HPA) axis is activated, which throws the automatic nervous system into disarray causing a rise in heart rate (Agelink et al., 2004; Juruena et al., 2018; Nederhof et al., 2015; Ulrich-Lai and Herman, 2009).
In this study, as a result of collecting heart rate data over seven days (including nights), we observed a significant difference in heart rates between patients and healthy controls, although this difference did not remain in the age-balanced subpopulation.
In regards to skin temperature, Avery et al. (1999) reported that patients with depression have higher body temperatures at night than healthy controls, and that after recovery, patients’ temperatures decrease. Our results also suggest that around the time period of 7:00 pm–11:00 pm, patients’ skin temperature increases significantly, which means skin temperature is a possible indicator for depression.
For UV light exposure, we found only a small difference for UV light exposure between depressed patients and healthy controls. There are several studies that have reported that UV light exposure has an effect on people’s emotional state, and UV light exposure has been used in the treatment of depression (Veleva et al., 2018).
Additionally, UV light is almost completely absent from interior light sources, but makes up a large part of light from the sun. Therefore, we believe that UV light exposure can indicate when someone goes outside, and so we used it as one of our measures in this study. It is possible that the lack of a significant difference in UV light exposure is due to the fact that much of modern life takes place indoors these days, which causes UV light exposure to be low overall regardless of depressive state.
Next, using individual longitudinal patient data, we compared data taken from the wearable devices regarding the worsening and lessening of symptoms.
From those results, we observed a tendency for daytime step count, energy expenditure, and body motion to be high during times of less severe symptoms, which shows that when depression symptoms are less severe, patients are more active during the day.
For sleep time, we found that during more severe symptom periods, patients slept comparatively more during earlier nighttime hours, whereas during less severe periods, patients slept more during later nighttime hours. We believe these results show that patients with less severe symptoms are sleeping well late at night, and when compared with healthy controls, both groups maintain similarly high regular activity levels earlier at night.
Additionally, using machine learning, we created algorithms to evaluate the presence and severity of depression symptoms, and tested those algorithms’ accuracy. For evaluating the presence of depression, we achieved an accuracy value of over 0.7 using data measured over a three-day period.
When we increased the data collection period to seven days, we did not observe an increase in accuracy. On the other hand, when using machine learning to evaluate illness severity, we reached a correlation coefficient of 0.48 for three days of data, whereas we achieved a correlation coefficient of 0.61 when analyzing seven days of data.
Based on these results, we believe that data collection over three days is adequate for diagnosing, but longer periods are needed for estimating illness severity.
Until now, there have been few studies attempting to use machine learning to analyze multimodal data collected from wearable devices in order to evaluate mood disorders. As stated in the introduction, Valenza et al. (2013) and Cho et al. (2019) used machine learning to analyze wearable device data to predict mood states.
Bourla et al. (2018) looked at the usefulness of unipolar depression evaluations utilizing wearable devices, and found that features like HRV and body temperature are useful in diagnosing depressive episodes.
There are relatively many more studies so far that have evaluated mood disorders using biological data taken from sources other than wearable devices. Lee et al. (2018) conducted a meta-analysis of 26 studies that used a variety of predictors to create machine learning algorithms for estimating the effectiveness of depression treatments.
According to this meta-analysis, most studies used the following types of predictors: neuroimaging, phenomenological (e.g., psychometric, neurocognitive, anthropometric, sociodemographic, psychiatric history), genetic (e.g., single nucleotide polymorphisms [SNPs]), or a combination of the above.
The meta-analysis reported that the combined results of these studies produced a prediction accuracy of 0.82. In a separate study, Ramasubbu et al. (2016) attempted to differentiate between MDD patients and healthy controls using fMRI data and a support vector machine.
They reported that while they were able to discern between patients with the heaviest symptoms and healthy controls (accuracy 66%, p = 0.012 corrected), they were unable to discern between patients with heavy symptoms (accuracy 52%, p = 1.0 corrected) and those with mild to moderate symptoms (accuracy 58%, p = 1.0 corrected).
Thus, the majority of previous research using machine learning has focused on labor-intensive brain image studies and genetic studies that involve large amounts of data inspection. But the wristband-type wearable device used in our study allows data for evaluating the presence and severity of depression to be collected more easily, which may be beneficial in real world clinical practice.
Furthermore, there are few studies that have used machine learning to estimate depression severity. Jiang et al. (2016) predicted illness severity based on HAMD scores using machine learning to analyze magnetoencephalography (MEG) data, and they reported a correlation coefficient of r = 0.38–0.68.
However, since only 22 datasets were analyzed with machine learning in that study, it is possible that it lacks reliability and generalizability. Therefore, the results of our study, which included 236 datasets and estimates illness severity with a correlation coefficient of 0.61, may be meaningful.
We found that skin temperature-related features consistently showed the greatest contribution to the machine learning algorithm’s predictive ability throughout all the models we built. Following skin temperature, sleep time-related features had the next highest significance in making predictions.
In biostatistical tests, instead of skin temperature or sleep time, heart rate was shown to have significant differences between healthy and depressed samples most frequently. Therefore, it is interesting to find that skin temperature, sleep time, and the correlation between skin temperature and sleep time contribute more to machine learning predictions than heart rate.
One possible explanation is that the statistical tests that we employed in this study test the differences in mean and standard deviation of each feature individually. It is likely that the nonlinear combinatory relationships between skin temperature, sleep time, and depression state were not discovered in the statistical tests, but were uncovered by the machine learning model.
Regarding the relationship between body temperature and sleep, it is reported that body temperature shifts with sleep, and the relationship can be disturbed by depression (Avery et al., 1999; Elsenga and Van den Hoofdakker, 1988; Lorenz et al., 2019). Hasler et al. (2010) compared 18 MDD patients and 19 healthy controls, and found that a larger phase angle difference between midsleep and the core body temperature minimum was associated with greater depression.
One common drawback to machine learning models is that results are often not easily explained in biological or clinical terms. However, the predictive capabilities of machine learning models can still be extremely valuable.
Also, in this case, we averaged the feature importance of features across the 10-fold cross-validation to find the features that are important in all of the models that we built, and referenced these features to findings in previously published research works. The results show that the consistently important features have been collaborated by previous research, which further confirms the validity of the machine learning models.
reference link: https://www.ncbi.nlm.nih.gov/pmc/articles/PMC7005437/