In this paper, we focus on automated detection of schizophrenia and bipolar disorder. For this task, we describe a simple and efective classifier, i.e. convolutional nearest neighbor. It provides a data-driven and objective approach for the detection of schizophrenia and bipolar disorder based on heart rate variability time series. According to our results, our approach is able to distinguish whether the selected person belongs to the patient group with an accuracy of 85% and area under receiver-operator characteristic curve of 0.92.
1. Introduction tive and time consuming as it depends on the physicians own skills, mood and the ability to establish contact with Psychotic disorders are a group of psychiatric disorders the patient. There are questionnaire tools such as the that are characterized by a disruption of the ability to Positive and Negative Symptom Scale (PANSS) that endistinguish between the internal experience of the mind able for a certain degree of objectivity by giving a more and the external reality. The two most common psychotic closed structure to the psychiatric interview [8, 9]. A disorders are schizophrenia and bipolar disorder (BD) [1]. structured interview is not a potential full solution to The prevalence of schizophrenia as well as BD in the this problem because it does not take into account the Western population can be estimated at 1% each [2, 3]. context in which the patient lives and acts. Therefore it Due to the fact that most of the human population has may potentially lead to diagnostic and disease severity little to no access to efective mental healthcare [ 4] and assessment errors [10]. This situation urges for tools that due to the large impact that these diseases have on qual- would give a higher degree of objectivity and repeatability of life [5], it is important to search for new automated ity in the diagnostic process. solutions in the field of disease diagnostics and supervi- When searching for data-driven, objective methods sion. for the diagnosis of schizophrenia and BD, heart rate
The diagnostic criteria for schizophrenia and BD are variability (HRV) could be a potential biomarker that is well defined by the International Classification of Dis- easily obtainable with consumer grade wearable devices eases (ICD-11) and the Diagnostic and Statistical Manual that function as electrocardiographs such as the Polar of Mental Disorders (DSM-5) [6, 7]. These criteria enable H10 [11, 12, 13]. HRV can be interpreted as an indirect a trained physician to decide on the diagnosis based on information on the function of the centrally regulated the interview with the patient. This process can be subjec- autonomic nervous system (ANS) and its parasympathetITAT Workshop on Bioinformatics and Computational Biology, ic/sympathetic balance [14, 15]. There is a documented Tatranske Matliare, Slovakia, Sept 22-26, 2023 correlation between this ANS balance and the psychotic * Corresponding author. process severity, where a lack of proper autonomic reg† These authors contributed equally. ulation constitutes itself as lower HRV in the psychotic $ buza@biointelligence.hu (K. Buza) patient [16]. http://www.biointelligence.hu/ (K. Buza) Thanks to the advancements in machine learning and (K. 0K0s0i0ą-ż0e0k0)2;-07010101--06040512-9(5K1.6B-0u7z0a9);(0W00.0M-0a0s0a2r-c0z2y0k1);-6220 data processing methods, heart rate and HRV could po0000-0002-0215-4674 (P. Głomb); 0000-0002-9419-7988 tentially be integrated into the clinical diagnostic process. (P. Gorczyca); 0000-0002-8009-7118 (M. Piegza) Attempts have been made for the continuous assessment © 2023 Copyright for this paper by its authors. Use permitted under Creative Commons License of psychotic symptoms in schizophrenia through HR analCPWrEooUrckReshdoinpgs IhStpN:/c1e6u1r3-w-0s.o7r3g ACttEribUutRion W4.0oInrtekrnsahtioonpal (PCCroBYce4.0e).dings (CEUR-WS.org) ysis, although the assessment of HR and accelerometry 10.000 on average. This corresponds to approximately instead of HRV raises the question whether this is an 2-hour recordings of RR intervals. We made the data ANS or physical activity biomarker [17]. Attempts for publicly available with its documentation to which we continuous assessment of HRV in BD used low precision refer for further details [21]. photoplethysmography (PPG) sensors that are inferior to ECG devices [11, 18]. 2.2. Convolutional Nearest Neighbor
To avoid the beforementioned problems in this paper we focus on HRV and use ECG sensors. We propose an As mentioned previously, our data contains 60 persons automated, data-driven and therefore objective method in total. In both versions of the data, i.e., both in the that may assist the diagnosis of schizophrenia and bipo- raw and preprocessed data, each of the participants is lar disorder. Our approach is based on machine learning associated with a RR time series, the length of which techniques. In particular, we aim to recognize the afore- is 10 000 on average. Given these relatively long time mentioned diseases based on HRV time series. Thus, we series and the fact that the disease is reflected in relatively consider the diagnosis of schizophrenia and bipolar disor- short segments of the RR time series, we consider short der as a time series classification task. To the best of our segments of the long time series. The length of each knowledge, ours is the first work that uses convolutional short segment is 25. In particular, we extract overlapping nearest neighbor for the classification of HRV data in segments according to a moving window schema with a the context of the diagnosis of schizophrenia and bipolar step size of 5 and 1 for training and test data, respectively. disorder. For example, in case of a step size of 1, a RR time series with length of 9500 results in 9500 − 25 + 1 = 9451 segments, whereas in case of a step size 5, the number of 2. Materials and Methods resulting segments is ⌈(9500 − 25 + 1)/5⌉ = 1891. We use -nearest neighbor with = 25 and cosine 2.1. Data distance to assign a score to each of these segments. The For this study, we recruited1 30 adult patients diagnosed score of a segment corresponds to the ratio of its neighand hospitalized for schizophrenia or bipolar disorder in bors that belong to the treatment (positive) class. For the Psychiatric Department in Tarnowskie Góry, Poland. example, if 7 out of 10 nearest neighbors of a segment
We decided to consider schizophrenia and BD together belong to the treatment class, the score of the segment as treatment group for the classifier due to their similari- is 0.7. ties in genetic and neuroanatomical features as well as In order to classify a person, we consider all segments the fact that both are defined as clinical syndromes or of her/his time series. The final score associated with disorders without regard to their pathophysiology [19]. the evaluated person is the average of the scores associMoreover, HRV can be understood as a transdiagnostic ated with the segments. This score may directly be used biomarker of psychopathology which further justifies our to assess the likelihood of the disease for the particular decision [20]. The control group consisted of 30 adults person, while all the scores associated with a set of exwithout any current psychiatric condition. The youngest periment participants may be used to assess the quality persons in the treatment and the control were 20 and 24, of the model, for example, in terms of the area under the respectively, while the oldest in both groups were 69. receiver-operator characteristic (ROC) curve.
For data collection, we used a wearable device of high Whenever a clear decision is needed on whether the quality, particularly Polar H10. It collects ECG signals, model considers a particular person to belong to the detects R peaks in the ECG and calculates the time be- treatment group or control group (for example, in order tween consecutive R peaks resulting in RR time series, to calculate accuracy, i.e., the ratio of correctly classified i.e., each value in the RR time series corresponds to the persons), we use a decision threshold of 0.5. time between two consecutive R peaks (i.e., the length of As our approach incorporates a rolling window techan RR interval). The RR time series acquired by the device nique with overlapping windows, it resembles a 1is referred to as the raw RR time series in the remainder dimensional convolutional kernel. Therefore, we call of the paper. it convolutional nearest neighbor.
After removing artifacts, we obtain the preprocessed RR time series. The length of both raw and preprocessed RR 3. Results and Discussion time series varies between 7.000 and 13.000 with around
1Participants were informed and asked for written consent prior to person-out cross-validation protocol. That is: in each admission to the experiment. The experimental procedure in the round of the cross-validation, we considered all the present study received approval from the local Bioethics Committee. segments belonging to one of the persons as test data, Approval nr.: BNW/NWN/0052/KB1/135/I/22/23. whereas the segments of all the other persons were con- various numbers of nearest neighbors used for classifisidered as training data. cation and they achieved similar results, indicating the
Our classifier achieved an accuracy of 0.850 and 0.833 stability of convolutional nearest neighbor in this dousing the preprocessed and raw data, respectively. The main. corresponding receiver-operator characteristic curves are The distributions of participants’ scores predicted by shown in Fig. 1 and Fig. 2. The area under the receiver- our convolutional nearest neighbor classifier are shown operator characteristic curve (AUC) is 0.92 and 0.91. in Fig. 3 and Fig. 4 in the case of the preprocessed and raw
As one can see, the performance on the preprocessed RR data, respectively. In both cases, the distributions of data is slightly better (one more person is classified cor- persons in the treatment (positive) and control (negative) rectly). The fact that convolutional nearest neighbor is groups are clearly diferent. As one can see, 0.5 may serve able to achieve relatively high accuracy, even in the case as an appropriate threshold value that allows classifying of raw data, indicates that convolutional nearest neigh- most of the persons correctly. bor is a promising candidate in applications where (semiautomated) preprocessing is not feasible, but enough 3.1. Diagnosis using less data data is available. Furthermore, we mention that we experimented with other versions of convolutional nearest Next, we simulate the situation in which the ECG of neighbor, in particular with various other distance met- the patient is observed for a shorter period of time and rics (Euclidean, Manhattan, dynamic time warping) and examine the performance of our classifier in this case. 3.2. Other classifiers
dynamic time warping, see e.g. [22] for an introductory survey. In contrast, recent approaches are based on deep learning [23] or the combination of deep learning techniques and dynamic time warping, see e.g. [24].
We run an initial experiment with such classifiers. As for dynamic time warping, we tried to use it as a distance measure instead of cosine distance within our convolutional nearest neighbor classifier. As expected, the resulting approach was orders of magnitude slower (i.e., more expensive computationally). However, the accuracy was roughly the same as in the case of other distances, such as cosine, Manhattan and Euclidean. Our finding is in line with the observations reported by Ding et al. [25], according to which the performance of Euclidean distance converges to the performance of dynamic time warping with increasing size of training data. We note that in our case, the training data contained roughly 100.000 segments in each cross-validation round which may explain why similar results were achieved with various distance metrics.
Regarding the approaches based on deep learning, fully convolutional neural networks (FCNs) were reported to be a strong baseline [23]. According to our initial observations, it was dificult to find appropriate hyperparameters (such as learning rate, batch size, etc.) for training FCNs and their performance on the test data was unstable.
Last, but not least, we mention that our AUC of 0.92 is slightly higher than the AUC achieved by Reinersten et al. [17], in case of two-day-long signals which is substantially longer than the RR time series we considered. Furthermore, in terms of AUC, our results seem to be competitive with many other approaches from the literature as well [26]. Nevertheless, we emphasize that both Reinersten et al. [17] and the works surveyed by Montazeri et al. [26] used diferent datasets.
consider the (i) first 100 and (ii) first 200 segments of the 4. Conclusion RR time series. This corresponds to a few minutes of RR time series and roughly 1% and 2% of the entire RR signal In this paper, we focused on the automated, data-driven we obtained for the given person. and therefore objective detection of schizophrenia and
Fig. 5 and Fig. 6 show the AUC in case of preprocessed bipolar disorder based on heart rate variability (HRV) and raw RR time series when using the first 100, first time series. We presented convolutional nearest neigh200 and all the segments. As expected, when using first bor, a simple, but efective approach for this task. We 100 or 200 segments only, the classifier is less accurate provided a detailed analysis of the predictions under varcompared with using the entire signal. Nevertheless, the ious conditions, such as raw and preprocessed RR signals. diference is marginal. Therefore the results show that We point out that, according to our observations, conaccurate classification may be achieved even if the ECG volutional nearest neighbor is a robust classifier w.r.t. is observed for a relatively short time. the settings of its hyperparameters, such as the distance metric or the number of nearest neighbors. Additionally, we showed that convolutional nearest neighbor is able to classify persons with a reasonable accuracy even if the HRV is only observed for a relatively short time of a few atric interview: Validity, structure, and subjectivminutes. ity, European Archives of Psychiatry and Clinical
As for our future work, we will examine convolutional Neuroscience 263 (2013) 353–364. doi:10.1007/
nearest neighbor more systematically under various con- s00406-012-0366-z.
ditions (e.g. diferent number of nearest neighbors, seg- [11] K. Hinde, G. White, N. Armstrong, Wearable
dement length) and we plan to perform more experiments vices suitable for monitoring twenty four hour heart
with neural networks. rate variability in military populations, Sensors 21
(
URL: http://www.nature.com/articles/nrdp201567. [14] dAo.iJ:1.0C.a1m0m07,/Ms.0M04a2li1k-,J0.1T9.-B0ig4g1e4r2,-G5..Breithardt, doi:10.1038/nrdp.2015.67. S. Cerutti, R. J. Cohen, P. Coumel, E. L. Fallen, H. L. [3] A. Jain, P. Mitra, Bipolar Disorder, 2023. URL: http: Kennedy, R. E. Kleiger, et al., Heart rate variability: //www.ncbi.nlm.nih.gov/pubmed/36579354. standards of measurement, physiological interpre[4] World Health Organization, Mental disorders, 2022. tation and clinical use. task force of the european URL: https://www.who.int/news-room/fact-sheets/ society of cardiology and the north american socidetail/mental-disorders. ety of pacing and electrophysiology, Circulation 93 [5] S. Evans, S. Banerjee, M. Leese, P. Huxley, The (1996) 1043–1065.
impact of mental illness on quality of life: A com- [15] G. G. Berntson, J. Thomas Bigger Jr, D. L. Eckberg, parison of severe mental illness, common mental P. Grossman, P. G. Kaufmann, M. Malik, H. N. Nadisorder and healthy population samples, Qual- garaja, S. W. Porges, J. P. Saul, P. H. Stone, et al., ity of Life Research 16 (2007) 17–29. doi:10.1007/ Heart rate variability: origins, methods, and inters11136-006-9002-6. pretive caveats, Psychophysiology 34 (1997) 623– [6] D. J. Stein, P. Szatmari, W. Gaebel, M. Berk, E. Vieta, 648.
M. Maj, Y. A. De Vries, A. M. Roest, P. De Jonge, [16] B. R. Benjamin, M. Valstad, T. Elvsåshagen,
A. Maercker, et al., Mental, behavioral and neu- E. G. Jönsson, T. Moberget, A. Winterton,
rodevelopmental disorders in the icd-11: an inter- M. Haram, M. C. Høegh, T. V. Lagerberg, N. E.
national perspective on key changes and controver- Steen, L. Larsen, O. A. Andreassen, L. T.
Westsies, BMC medicine 18 (2020) 1–24. lye, D. S. Quintana, Heart rate
variabil[7] American Psychiatric Association, Diagnostic and ity is associated with disease severity in
psystatistical manual of mental disorders: DSM-5, chosis spectrum disorders, Progress in
NeuroAmerican Psychiatric Association Washington, DC, Psychopharmacology and Biological Psychiatry 111
2013. (
FullText/284894. doi:10.1159/000284894. [18] A. Stautland, O. B. Fasmer, P. Jakobsen, B. Osnes, [9] N. C. Andreasen, The Scale for the Assess- J. Torresen, T. Nordgreen, K. J. Oedegaard, P178. ment of Negative Symptoms (SANS): Concep- heart rate variability as biomarker for bipolar matual and theoretical foundations, British Jour- nia, Biological Psychiatry 91 (2022) S159. nal of Psychiatry 155 (1989) 49–52. doi:10.1192/ [19] C. Dacquino, P. De Rossi, G. Spalletta, Schizophres0007125000291496. nia and bipolar disorder: the road from similarities [10] J. Nordgaard, L. A. Sass, J. Parnas, The psychi- and clinical heterogeneity to neurobiological types,