Externally Validated Deep Learning Model to Identify Prodromal Parkinson’s Disease from Electrocardiogram

doi:10.21203/rs.3.rs-2277951/v1

Download PDF

Article

Externally Validated Deep Learning Model to Identify Prodromal Parkinson’s Disease from Electrocardiogram

https://doi.org/10.21203/rs.3.rs-2277951/v1

This work is licensed under a CC BY 4.0 License

Journal Publication

published 29 Jul, 2023

Read the published version in Scientific Reports →

You are reading this latest preprint version

Little is known about Electrocardiogram (ECG) markers of Parkinson’s disease (PD) during the prodromal stage. The aim of the study was to build a generalizable ECG-based fully automatic artificial intelligence (AI) model to predict PD risk during the prodromal stage, up to 5 years before incidence of the disease. This retrospective case-control study included samples from Loyola University Chicago (LUC) and University of Tennessee-Methodist Le Bonheur Healthcare (MLH). Cases and controls were matched according to specific characteristics (date, age, sex and race). Only data available at least 6 months before PD diagnosis was used as the model’s input. Data from LUC spanned back to May 2014 while that from MLH spanned to January 2015. PD was denoted by at least two primary ICD diagnostic codes, namely ICD9 332.0, ICD10 G20. PD incidence date was defined as the earliest of first PD diagnostic code or PD-related medication prescription. Prediction of prodromal PD (6-months to 5-years preceding PD diagnosis) was the primary outcome of this research. Three time windows were set: 6 months-1year, 6months-3 years and 6months – 5 years. A novel deep neural network using standard 10-second 12-lead ECG was used to predict PD risk at the prodromal phase. This model was compared to multiple feature engineering-based models. Subgroup analyses for gender, race and age were also performed. A one-dimensional convolutional neural network (1D-CNN) was used to predict PD risk (or identify prodromal PD) from standard 10 second 12-lead ECGs collected between 6 months to 5 years before a clinical diagnosis. The prediction model was built using MLH data and externally validated on LUC data. 131 cases/1058 controls at MLH and 29 cases/165 controls at LUC were identified. The model was trained on 90% of the MLH data, internally validated on the remaining 10% and externally validated on LUC data. The best performing model resulted in an external validation of AUC = 0.67 when predicting prodromal PD at any time between 6 months and 5 years. The accuracy increased when using ECGs to predict prodromal PD within 6 months to 3 years, with an external validation AUC of 0.69 and achieving highest AUC when predicting PD within 1 year before onset (AUC of 0.74). A predictive model that can correctly classify individuals with prodromal PD was developed using only raw ECGs as inputs. The model was effective in predicting prodromal PD within an independent cohort, particularly closer to disease diagnosis. The ECG-based model outperformed multiple models built using ECG feature engineering. Subgroup analyses showed that some subgroups, including females and those of over 60 years of age, might benefit from closer monitoring, especially when symptoms start becoming more evident but not enough to make a diagnosis. This research highlights that standard ECGs may help identify individuals with prodromal PD for cost-effective early detection and inclusion in disease-modifying therapeutic trials.

Health sciences/Neurology/Neurological disorders/Parkinsons disease

Physical sciences/Mathematics and computing/Computer science

Health sciences/Biomarkers/Predictive markers

Parkinson’s Disease (PD) is a systemic disease that is currently diagnosed when classic motor symptoms, most notably tremors, become evident and generally manifest after 50% of substantia nigra dopaminergic neurons are already dead or dying and with extensive striatal dopaminergic deafferentation^1–4. However, pathologic changes in the lower brainstem and peripheral autonomic nervous system likely begin years or even decades before significant nigral damage occurs^5,6, and symptoms referable to these sites of early pathologic injury may manifest many years before motor PD is diagnosed³. Interventions that delay or slow down the progression from prodromal disease to parkinsonism could be implemented during this pathologic evolution if PD patients could be identified early with confidence.

Lewy pathology is found throughout the autonomic nervous system in PD.⁷ Changes in the pathology of the sympathetic and parasympathetic ganglia, cardiac nerves and cardiac deafferentation are consistently seen in early PD ^8–11 and post-mortem examinations associate these pathologies with incidental nigral Lewy bodies (ILB).¹⁰ For this reason, cardiac sympathetic deafferentation as measured by ¹²³I-Metaiodobenzylguanidine (¹²³I-MIBG) scintigraphy is designated as a supportive criterion for the clinical diagnosis of PD in the MDS-PD diagnostic criteria.¹² However, MIBG scintigraphy is invasive and expensive, and is not a viable tool for population-level screening for PD risk.

Sympathetic and parasympathetic nerve inputs mediate adaptive responses to varying physiological conditions and cardiovascular demands, all of which can manifest as heart rate variability (HRV). HRV can be assessed by measuring the duration between consecutive R waves on an ECG.¹³ HRV is reduced in PD, likely reflecting underlying cardiac autonomic Lewy pathology and deafferentation^14–22. In addition, HRV has recently been shown to correlate with striatal dopamine depletion in PD.²³ Paralleling MIBG observations, evidence suggests that HRV may be reduced in prodromal PD. HRV is reduced in rapid eye movement (REM) sleep behavior disorder (RBD), ^24,25 a condition that is thought to be a biomarker and strongest predictor for prodromal PD, due to its direct correlation to the progression to PD ^26,27.

A previous study using data from Atherosclerosis Risk in Communities Study Description (ARIC), a large prospective study, found that low HRV was associated with 2-3-fold increased subsequent risk of PD over a mean of 18 years follow up period.²⁸ In addition, there has been research into predicting PD or prodromal PD using numerous clinical risk factors in combination with biological components extracted from serum samples, such as cytokines and chemokines within a decision tree algorithm²⁹. Research by Akbilgic et al. (2022) used cardiac electrical activity signal information from 60 subjects within a logistic regression model following a Probabilistic Symbolic Pattern Recognition (PSPR) method from the Honolulu Asia Aging Study (HAAS) to identify patients at high risk of PD but without a pre-specified time window³⁰. While this research reported an average AUC of 0.835, the HAAS cohort was a small data sample and not representative of the general population. While HRV metrics, typically measured using 5-minute ECGs, have been used in multiple studies, there is some debate whether such classical metrics can indeed capture substantial information on the likelihood of development of PD³⁰. In previous research, models were developed to predict future PD using machine learning algorithms that incorporate multiple features derived from the full waveform of standard 10-second ECGs.³⁰ Such research indeed provides substantial indication that ECG-based signals can be used as a predictor of PD, with possibility of predicting PD onset at a timely manner before diagnosis.

Other research, such as that by Yuan et al. (2021), compared artificial intelligence approaches using demographic and clinical diagnoses as features within a logistic regression framework as well as a deep recurrent neural network (RNN) which considered sequences of patient interactions with the health care system as inputs to predict PD within multiple time window, with a maximum window of 720 days before PD diagnosis. The highest accuracy, as expected, was achieved very close to the time of PD onset, reducing in accuracy of approximately AUC = 0.67 and AUC = 0.71 at 720 days prior to PD onset using logistic regression and RNN, respectively³¹. While informative, the extraction of such data can be difficult and might not be available at all clinical institutes. Research by Karabayir et al.³² (2022) showed that prodromal PD can be predicted with certain accuracy using simple clinical variables, including age, caffeine intake, body mass index, smoking status as well as multiple screening/diagnosis tests. This research showed that identifying PD can be difficult especially since a real diagnosis is generally related to Lewy Body density, which can accurately be measured during autopsy. The increase in Lewy body density during progression to PD onset has been directly related to sympathetic nerve damage within the heart as well as cardiac autonomic dysfunction, the signals of which have the potential to act as a proxy for ECG-based detection of PD and its prodromal stage^25,33.

This research builds on prior artificial intelligence (AI) approaches but uses raw waveform ECGs rather than features extracted from ECGs. There is substantial utility, generalizability and simplicity of raw 12-lead ECGs, which are easily and routinely collected, to predict prodromal (and risk of) PD within a specified time window. This study uses two independent study populations with available electronic ECGs obtained prior to PD onset, with one dataset used for model building and the other used for external validation.

Our initial EHR-based search identified 131 eligible PD cases and 1,058 controls from the MLH database, with a total of 428 and 3584 available ECGs, respectively. After validation by manual chart review of the LUC dataset, this dataset included 29 PD cases and 1 The model’s training process was stopped once reaching consecutive decrease (five times in a row) on the AUC of MLH internal validation dataset. age-gender matched controls with one ECG each. The baseline characteristics of these cohorts are summarized in Table 1.

Table 1

Baseline Characteristics
	MLH (N = 1189 with 131 cases)			LUC (N = 194 with 29 cases)
	Total (n = 1189)	Case (n = 131)	Control (n = 1058)	Total (n = 194)	Case (n = 29)	Control (n = 165)
Mean Age at ECG (SD)	70.1 (9.5)	72.6 (8.0)	69.8 (9.6)	71.2 (8.7)	69.8 (8.7)	71.4 (8.8)
Mean Age at PD diagnosis (SD)	75.3 (7.7)	75.3 (7.9)	-	72.3 (8.0)	71.0 (8.8)	-
Mean Years between ECG and PD diagnosis (SD)	2.7 (1.9)	2.7 (1.9)	-	1.7 (1.0)	1.65 (1.0)	-
Male n (%)	639 (53.7%)	77 (58.8%)	562 (53.1%)	119(61.3%)	18 (62.1%)	101 (61.2%)
Race n (%)
White	753 (63.3%)	86 (65.7%)	667(63.0%)	149(76.8%)	23(79.3)	126(76.4)
Black or African American	412 (34.7%)	41 (31.3%)	371(35.1%)	22(11.3%)	3(10.4)	19(11.5)
Multiple	10 (0.8%)	0(0%)	10(1.0%)	0(0%)	0(0%)	0(0%)
Other/Unknown	5 (0.4%)	2(1.5%)	3(0.3%)	19(9.8%)	3(10.4)	16(9.7)
Asian	8 (0.7%)	2(1.5%)	6(0.6%)	4(2.1%)	0(0.0%)	4(2.4%)
Amer. Ind/Alaska Native	1 (0.1%)	0 (0%)	1 (0.1%)	0 (0%)	0 (0%)	0 (0%)

The developed 1D-CNN model, as visualized in Fig. 2, used 90% MLH training data and preliminary tested on the 10% MLH as the internal validation data. Among the epochs of training, a maximum AUC of 0.73 with CI of 0.51–0.96 was obtained when tested on the 10% MLH internal validation, with a sensitivity of 0.43 and a specificity of 0.96 when predicting prodromal PD up to 5 years before disease onset. This model, having the highest AUC from all epochs, was used as the final model for external validation of PD onset within 5 years. The accuracy increased slightly when predicting prodromal PD within 3 years of onset, with AUC = 0.69 (0.57–0.82), with the highest accuracy achieved, as expected, when predicting risk of onset up to 1 year of diagnosis with an AUC of 0.74 (0.62–0.87). The developed models for each time period (6 months to 5 years before PD onset) were validated using LUC data (Table 2). However, preference was given to the model predicting prodromal PD within 5 years, acknowledging that this achieved moderate AUC. When this model was validated on the LUC data, the model resulted in an AUC of 0.67 (0.54,0.79), correctly classifying 12 PD cases from 29 real PD cases at the prodromal phase, and 123 controls among 165 controls, with a specificity of 0.75 and sensitivity of 0.41.

Table 2

Sensitivity Analysis for the CNN model on LUC external validation
Subgroup		Time from ECG until PD incident diagnosis date
Subgroup		6 months − 1 year	6 months − 3 years	All (6 months − 5 years)
All subjects		0.74 [0.62–0.87]	0.69 [0.57–0.82]	0.67 [0.54–0.79]
Gender	Male	0.73 [0.57–0.9]	0.71 [0.55–0.87]	0.69 [0.53–0.85]
Gender	Female	0.79 [0.6–0.98]	0.66 [0.45–0.86]	0.63 [0.43–0.84]
Race	White	0.71 [0.56–0.86]	0.70 [0.56–0.85]	0.67 [0.52–0.81]
Race	Black	0.65 [0.27–1.03]	0.65 [0.27–1.03]	0.65 [0.27–1.03]
Age	>=60	0.73 [0.6–0.87]	0.70 [0.57–0.83]	0.67 [0.53–0.80]
Age	< 60	0.66 [0.26–1.05]	0.66 [0.26–1.05]	0.66 [0.26–1.05]

In addition to the 1D-CNN, models using multiple feature-engineering methods, including descriptive statistics extracted from 12 leads of the ECGs, sample entropy, PSPR etc., were also developed to predict prodromal PD within a 5-year time period. This was done to compare multiple models with a raw ECG-based model. From all the methods listed in Supplementary Table S1, the highest LUC validation accuracy achieved was AUC = 0.54 when using Fourier transformation and Discrete wavelet transformation. When all the different features extracted from all the different methods were included within one main LightGBM model, the AUC increased to 0.61 (0.52–0.79). These results show that the external validation accuracy obtained from deep learning on raw digital ECGs was superior to those obtained from machine learning approaches using engineered ECG features (Supplementary Table S1). Since there was no major improvement in the LightGBM model utilizing all the extracted features, the 1D-CNN ECG model was the basis of the subgroup analyses.

Subgroup Analysis. Although statistical power was limited for some analyses, we implemented a comprehensive subgroup analysis of the LUC external validation cohort, stratifying by gender, race, and age. We also explored how predictive accuracy varied by time from ECG until PD incidence date (Table 2).

As shown in Table 2, prediction accuracy of the model improved within a shorter time interval between the ECG and the PD diagnosis. The 1D-CNN model predicted prodromal PD in males at the highest overall accuracy, with AUC = 0.69 (0.53–0.85), compared to any other subgroup within a 5-year and 3-year prodromal period with very little difference between accuracy for gender, race and age subgroups for both time periods. The noticeable difference in accuracy is when predicting PD within 1 year of onset, with this model predicting risk at moderate-high accuracy for females (AUC = 0.79 (0.60–0.98)) followed by people, irrespective of gender and race, over or equal to 60 years of age (AUC = 0.73 (0.60–0.87)).

While there has been an increase in studies researching early identification of PD through biological models, there still remains a major gap in early detection and, even more so, prediction of PD within the prodromal stage³¹. While multiple studies have used other clinical diagnoses as precursors to predict PD, this might limit the use of such models due to the possible lack of symptoms at the pre-screening stage. Therefore, accurate identification of prodromal PD is essential for implementation of cost-effective clinical trials of putative neuroprotective agents, and ultimately for risk stratification and population-level disease prevention once therapeutic efficacy is established for one or more agents. In order to be effective on a population level, a predictive test of risk of PD should be universally available, cost-effective and non-invasive. Studies found that heart rate variability (HRV) determined from 5-minute ECGs is reduced in prevalent PD, and results from a single prospective study showed that lower HRV was associated with an increased risk of incident PD.³⁴ Prior work by our group utilized machine learning approaches to predict prodromal PD using clinical variables in one study³² and a proof-of-concept study using standard 10-second printed ECGs in another study³⁰. Both studies resulted in moderate accuracy, however the data and model developed in the latter study was biased towards an elderly population and also lacked an external cohort for model validation.³⁰

The current research builds on the use of ECGs in Parkinsonism studies, using a deep learning approach to develop a predictive model for prodromal PD using standard 10-second raw digitized ECGs stored in a large healthcare system EHR and validate this model using an independent cohort from a different healthcare system EHR. The developed model had moderately good classification accuracy in the validation cohort, with an AUC of 0.74 between 6 and 12 months before PD diagnosis and an AUC of 0.67 when PD 5 years before diagnosed onset. While AUC can be deemed as somewhat low, it should be noted that the main aim of this model was to help clinicians in assessing probable risk of people developing the disease and not for diagnosis. This means that the model’s ability to correctly distinguish between the majority of cases and controls, collectively, has important implications for the potential in its use for early detection and pre-screening.

Furthermore, we validated PD diagnosis and incidence dates in the external LUC validation population thus ensuring accurate performance metrics. Given that the model was developed in a cohort without chart reviews in MLH, a proportion of which were likely mislabeled, the predictive performance of the model in the replication cohort provides an important indication for the use of ECGs in prodromal PD prediction as well as for future studies. This is even more important since almost half of the cases were correctly identified as such, and most of the controls were also classified properly. We acknowledge that the developed deep learning model needs improvement in case identification, however, it should be noted that the MLH dataset used for training was largely imbalanced between cases and controls, making the model slightly more effective in identifying controls. The model’s performance can be improved following re-training using a larger cohort or the implementation of transfer learning in combination with a larger cohort.

In comparison to the machine learning models utilizing ECG feature-engineering as inputs (Supplementary Table S1), the CNN model predicted prodromal PD at higher accuracy. These results are also comparable to resulted in published research that included multiple different biological and demographic information, some of which can be quite costly to obtain (e.g. serum analysis), within a machine learning system. Furthermore, we emphasize the fact that while the AUC results of the CNN model for the prediction of prodromal PD within 5 years before onset are somewhat moderate, this research shows that PD cardiac markers, which are found to be present before any of motor changes involved in the onset of PD³⁵, can act as signals within an artificial intelligence framework without the need of full ECG feature extraction. Furthermore, there is the potential to include demographic and clinical variables as additional inputs with the predicted outcomes from the ECG CNN model, to increase the accuracy of predicting prodromal PD within an extended time window. Karabayir et al.³² (2022) showed that there is added benefit in including demographic and clinical variables in the prediction of prodromal PD, allowing for opportunity for exploration of improving this work. Such variables can also extend to the inclusion of data from the Cognitive Abilities Screening Instrument (CASI), olfactory tests and the simple choice reaction times test. At the time of development of the models in this research, such clinical and demographic variables were not available from the MLH and LUC data sources.

Overall, the CNN model developed in this research classified individuals based solely on a 10-second ECG input to an artificial intelligence framework. It is highly likely that incorporation of our ECG-based predictive model with other known prodromal markers of PD would result in improved risk stratification for PD. This further suggests the importance of routine collection of ECGs within cohort studies aiming to understand pathophysiology of PD and other neurodegenerative diseases.

This study has some limitations. Predominantly, we were able to perform chart reviews on LUC data, but not on MLH data. Based on the high false positive rate of PD diagnosis based exclusively on EHR ICD codes at LUC, there might have been the potential that a portion of MLH data included false case annotations, which would have negatively influenced model training and consequently its testing. Therefore, implementing transfer learning on our final deep learning model and retraining it on large well-annotated datasets can result in a more accurate PD prediction tool.

Another limitation of this study was that our chart reviews at LUC showed that out of 47 cases identified in LUC, only 29 of them were real PD cases. Therefore, our study also highlights the challenges around finding PD patients via her-based queries. However, it is worth noting that the developed PD risk prediction model produced significantly higher risk prediction for validated cases compared to controls.

We, note that the raw ECG data used in this study was gathered from two different cardiology information systems: GE MUSE at LUC and Epiphany at MLH. Therefore, a major strength of this study lies in the fact that the concordant results between internal testing and the validation process provides evidence that raw ECGs exported from different platforms can be used reliably within AI frameworks, with minimal pre-processing.

Future works largely lie in the inclusion of demographic, clinical variables and results from known PD diagnoses tests within the current deep learning model. Continuous improvement of the models’ prediction accuracy, especially following the use of larger cohort data, and inclusion of other ‘simple’ variables can lead to the integration of these models within ECG-capable smart wearables. The development of smart applications within wearables allows for cost-effective, non-invasive and minimal clinical burden for screening people for PD risk, allowing for early detection, quicker follow-up times and timely intervention strategies.

The result from this research shows that simple 10-second ECGs allow for the development of a deep learning predictive model that correctly classifies individuals with prodromal PD with modest accuracy. This model was effective in distinguishing between PD cases and controls within an independent cohort, increasing in accuracy closer to disease diagnosis, although still notably within the prodromal stage of the disease. Therefore, the use of standard ECGs, which are easily and routinely collected by healthcare institutes, may help identify individuals at high-risk of PD, allowing for timely inclusion in disease-modifying therapeutic trials and possible intervention strategies to help delay or slow down disease progression.

Standard protocol approvals and patient consents. This retrospective case-control study was conducted using data from two hospital centers in the United States (Loyola University Chicago (LUC), Maywood, IL and University of Tennessee Health Science Center (UTHSC), Memphis, TN). The study was reviewed and approved by the institutional review boards of the participating centers (Loyola University Chicago and University of Tennessee Health Science Center), and all methods were carried out in accordance with the relevant guidelines and regulations. Due to its retrospective nature and exempt status, informed consent forms from study participants waived off by the institutional review boards of the participating centers.

Study Cohorts. This study used electronic health record (EHR)-based datasets from Loyola University Chicago (LUC), Chicago, IL and University of Tennessee-Methodist Le Bonheur Healthcare (MLH), Memphis, TN. The study design is summarized in Fig. 1. The participants in the analytic cohort were aged 26–89 years at the time that the ECG was taken.

Case Identification. In both datasets, we identified patients aged 26–89 years who had received at least two primary ICD diagnostic codes for PD (ICD9 332.0, ICD10 G20) spanning research datasets dating back to May 2014 at LUC and January 2015 at MLH. PD incidence date was defined as the earliest of the first PD diagnostic code or PD-related medication prescription. For the external validation phase, we included individuals from LUC who had a standard 12-lead ECG within 5 years of the incidence date, which this was the longest time interval we had at the time of the study. However, to have a larger training sample, we included individuals from MLH within 10-years prior to the incidence date.

Case Validation. We reviewed the medical records of included individuals from the LUC dataset to confirm the PD diagnosis and date of disease incident diagnosis. To carry out this evaluation on the LUC data, the patients’ charts were accessed using the Epic software, and the related clinical notes were reviewed by a movement disorders specialist (co-author KC). ‘Care Everywhere’ notes in Epic from other healthcare systems were also examined when available. We considered diagnostic consistency, clinical features, response to medication, and possible alternative diagnoses. All patient charts were also searched for the presence of prodromal features and their time of onset, date of ECG tracing examined, and the presence of medical diagnoses or medications that could potentially confound or alter the results of the ECG tracing.

Control identification. Randomly selected control subjects were matched to cases by age at ECG, date of ECG recording, race and sex, have been in the system for at least 10 years (for MLH) or 5 years (for LUC) and had no diagnostic codes for PD or parkinsonism. Considering the available resources to extract raw digital ECGs from cardio servers, 131 eligible cases and 1058 controls from MLH, and 29 eligible chart-reviewed cases (47 cases before manual chart review) and 165 controls from LUC were retrieved as the final cohort sample.

Covariate data. Covariates were also obtained from the EHR including age, gender, and race/ethnicity.

ECG Data. Digital ECG, referring to time-voltage data, was obtained for each case and control. The ECGs at LUC were exported from MUSE Cardiology Information System in XML format at 500Hz. Base 64 encoded waveform data was further decoded using the Python programming language. ECGs at MLH were exported in DICOM format and the associated waveforms were parsed out. Some MLH ECGs were recorded at 500Hz and some at 250 Hz and therefore, to ensure comparability and based on our previous successful application on utilization of ECG via deep learning,^36,37 all ECGs at 500 Hz were down-sampled to 250Hz by removing one of every two amplitudes from each lead. In addition, we excluded the first second of each ECG before inputting them into the model. This was done to avoid any possible noise due to the patient’s movement at the beginning of the ECG recording.

Method Development and External Validation. Because the MLH study cohort was significantly larger than the LUC cohort, we used MLH data for training and LUC data for external validation. We note that we had access to one ECG per patient from LUC. However, we could retrieve multiple ECGs per patient at MLH, and we utilized all ECGs during training. We split the MLH dataset into 90% training and 10% internal validation sets based on patient rather than ECGs since some patients have multiple ECG. All available ECGs were used for the 90% training dataset. However, while multiple ECGs were available, in the 10% MLH internal validation dataset only one randomly selected ECG per patient was used. This was done with the aim of achieving representative samples to that within the LUC external validation dataset. The model’s hyperparameters were tuned based on the highest accuracy when using the internal validation set. This model, i.e. the one providing the highest AUC, was used for external validation on LUC data. In addition to the AUC, the best performing model was assessed using sensitivity, specificity, positive predictive value, and negative predictive value.

Deep Learning Method. We used a one-dimensional convolutional neural network (1D-CNN) with raw digital ECGs as inputs with the risk for PD being the determined output. The 1D-CNN, inspired from ResNet,³⁸ takes advantage of skipping connections between layers, which eases the optimization of the kernels in the network and therefore obtains robust and more generalizable models than those yielded by traditional or shallow networks. CNN architectures include several hyperparameters including the number of convolutional filters, the kernel dimension, and the stride parameter of convolution layers. The parameters and the overall architecture of the CNN are depicted in Fig. 2. Furthermore, we utilized the leaky rectified linear units (Leaky-RELU) activation function³⁹ to obtain a better representation of the ECG by activating the intermediate neurons and allowing negative inputs. After the hyperparameter tuning process, dropout layers with a rate of 0.1 were included within the CNN architecture to penalize the kernels and help prevent model overfitting⁴⁰. The Adam optimization algorithm⁴¹ (beta_1 = 0.9, beta_2 = 0.999 and learning rate = 0.001) was used to optimize the kernels in the model. The batch size for data input was set to 128. The model’s training process was stopped once it reached consecutive decrease (five times in a row) on the AUC of MLH internal validation dataset.

Conventional Machine Learning Method. To compare the proposed deep learning algorithm, we also utilized Light Gradient Boosting Machine (LightGBM) by inputting ECG features, which were extracted from a variety of signal processing algorithms. ⁴² In the development of the LigthGBM model, the same five-fold cross validation strategy used in the development of the CNN model was used to ensure comparability. Details of the feature engineering-based machine learning model can be found in Supplementary Material.

Data Availability

Requests for patient-related data not included in the article will not be considered. The code built in can be available from the corresponding author upon request.

Author contributions statement

Types of authors’ roles: 1. Research project: A. Conception, B. Organization, C. Execution; 2. Data A. Preparing and/or handling, B. Chart Reviews 3. Statistical and Machine Learning Analyses: A. Design, B. Execution, C. Review and Critique; 4. Manuscript: A. Writing of the First draft, B. Review and Critique. IK contributed to 2A, 3A, 3B, 3C, 4A, and 4B. FG contributed to 3A, 3B, 3C, 4A, and 4B. SMG contributed to 1A, 1B, 1C, 1D, 4A, and 4B. RK and RLD contributed to 1A, 1B, 1C, 1D, and 4B. KC contributed to 2B and 4B. LC contributed to 2A and 4B. JLJ, KB, WR, HP, KM, LB and CT contributed to 4B. OA contributed to 1A, 1B, 1C, 1D, 3A, 3C, 4A, and 4B.

Competing interests

The authors declare no competing interests.

Additional information

We thank Michael J Fox Foundation for supporting this research (MJFF Grant ID 17267, PI Akbilgic). The funding sources had no role in the design and conduct of the study; collection, management, analysis, and interpretation of the data; preparation, review, or approval of the manuscript; and decision to submit the manuscript for publication.

Scherman D, Desnos C, Darchen F, Pollak P, Javoy-Agid F, Agid Y. Striatal dopamine deficiency in Parkinson's disease: role of aging. Research Support, Non-U.S. Gov't. Ann Neurol. Oct 1989;26(4):551–7. doi:10.1002/ana.410260409
Bernheimer H, Birkmayer W, Hornykiewicz O, Jellinger K, Seitelberger F. Brain dopamine and the syndromes of Parkinson and Huntington. Clinical, morphological and neurochemical correlations. J Neurol Sci. Dec 1973;20(4):415–55.
Poewe W, Seppi K, Tanner CM, et al. Parkinson disease. Nat Rev Dis Primers. Mar 23 2017;3:17013. doi:10.1038/nrdp.2017.13
Obeso JA, Stamelou M, Goetz CG, et al. Past, present, and future of Parkinson's disease: A special essay on the 200th Anniversary of the Shaking Palsy. Mov Disord. Sep 2017;32(9):1264–1310. doi:10.1002/mds.27115
Braak H, Del Tredici K, Rub U, de Vos RA, Jansen Steur EN, Braak E. Staging of brain pathology related to sporadic Parkinson's disease. Neurobiol Aging. Mar-Apr 2003;24(2):197–211. doi:S0197458002000659 [pii]
Braak H, de Vos RA, Bohl J, Del Tredici K. Gastric alpha-synuclein immunoreactive inclusions in Meissner's and Auerbach's plexuses in cases staged for Parkinson's disease-related brain pathology. Neurosci Lett. Mar 20 2006;396(1):67–72.
Dickson DW, Fujishiro H, Orr C, et al. Neuropathology of non-motor features of Parkinson disease. Parkinsonism Relat Disord. Dec 2009;15 Suppl 3:S1-5. doi:10.1016/S1353-8020(09)70769-2
Iwanaga K, Wakabayashi K, Yoshimoto M, et al. Lewy body-type degeneration in cardiac plexus in Parkinson's and incidental Lewy body diseases. Neurology. Apr 12 1999;52(6):1269–71.
Orimo S, Takahashi A, Uchihara T, et al. Degeneration of cardiac sympathetic nerve begins in the early disease process of Parkinson's disease. Brain Pathol. Jan 2007;17(1):24–30. doi:BPA032 [pii] 1111/j.1750-3639.2006.00032.x
Orimo S, Uchihara T, Nakamura A, et al. Axonal alpha-synuclein aggregates herald centripetal degeneration of cardiac sympathetic nerve in Parkinson's disease. Brain. Mar 2008;131(Pt 3):642–50.
Fujishiro H, Frigerio R, Burnett M, et al. Cardiac sympathetic denervation correlates with clinical and pathologic stages of Parkinson's disease. Mov Disord. Jun 15 2008;23(8):1085–92.
Postuma RB, Berg D, Stern M, et al. MDS clinical diagnostic criteria for Parkinson's disease. Mov Disord. Oct 2015;30(12):1591–601. doi:10.1002/mds.26424
ESC/NASPE. Heart rate variability. Standards of measurement, physiological interpretation, and clinical use. Task Force of the European Society of Cardiology and the North American Society of Pacing and Electrophysiology. Eur Heart J. Mar 1996;17(3):354–81.
Haapaniemi TH, Pursiainen V, Korpelainen JT, Huikuri HV, Sotaniemi KA, Myllyla VV. Ambulatory ECG and analysis of heart rate variability in Parkinson's disease. J Neurol Neurosurg Psychiatry. Mar 2001;70(3):305–10.
Kallio M, Suominen K, Bianchi AM, et al. Comparison of heart rate variability analysis methods in patients with Parkinson's disease. Med Biol Eng Comput. Jul 2002;40(4):408–14.
Maetzler W, Karam M, Berger MF, et al. Time- and frequency-domain parameters of heart rate variability and sympathetic skin response in Parkinson's disease. J Neural Transm. Mar 2015;122(3):419–25. doi:10.1007/s00702-014-1276-1
Arnao V, Cinturino A, Mastrilli S, et al. Impaired circadian heart rate variability in Parkinson's disease: a time-domain analysis in ambulatory setting. BMC Neurol. Apr 23 2020;20(1):152. doi:10.1186/s12883-020-01722-3
Bugalho P, Mendonça M, Lampreia T, Miguel R, Barbosa R, Salavisa M. Heart rate variability in Parkinson disease and idiopathic REM sleep behavior disorder. Clin Auton Res. Dec 2018;28(6):557–564. doi:10.1007/s10286-018-0557-4
Cygankiewicz I, Zareba W. Heart rate variability. Handb Clin Neurol. 2013;117:379–93. doi:10.1016/b978-0-444-53491-0.00031-6
Heimrich KG, Lehmann T, Schlattmann P, Prell T. Heart Rate Variability Analyses in Parkinson's Disease: A Systematic Review and Meta-Analysis. Brain Sci. Jul 21 2021;11(8)doi:10.3390/brainsci11080959
Li Y, Wang J, Li X, Jing W, Omorodion I, Liu L. Association Between Heart Rate Variability and Parkinson's Disease: A Meta-analysis. Curr Pharm Des. 2021;27(17):2056–2067. doi:10.2174/1871527319666200905122222
Terroba-Chambi C, Abulafia C, Vigo DE, Merello M. Heart Rate Variability and Mild Cognitive Impairment in Parkinson's Disease. Mov Disord. Dec 2020;35(12):2354–2355. doi:10.1002/mds.28234
Kitagawa T, Umehara T, Oka H, et al. Association between heart rate variability and striatal dopamine depletion in Parkinson's disease. J Neural Transm (Vienna). Dec 2021;128(12):1835–1840. doi:10.1007/s00702-021-02418-9
Valappil RA J, Black J, Broderick MJ, et al. Assessment of heart rate variability during wakefulness in patients with RBD. Abstract. Movement Disorders. 2009;24(Suppl. 1):S321-S322.
Postuma RB, Lanfranchi PA, Blais H, Gagnon JF, Montplaisir JY. Cardiac autonomic dysfunction in idiopathic REM sleep behavior disorder. Mov Disord. Oct 30 2010;25(14):2304–10. doi:10.1002/mds.23347
Postuma RB, Gagnon JF, Bertrand JA, Genier Marchand D, Montplaisir JY. Parkinson risk in idiopathic REM sleep behavior disorder: preparing for neuroprotective trials. Neurology. Mar 17 2015;84(11):1104–13. doi:10.1212/WNL.0000000000001364
Schenck CH, Boeve BF, Mahowald MW. Delayed emergence of a parkinsonian disorder or dementia in 81% of older men initially diagnosed with idiopathic rapid eye movement sleep behavior disorder: a 16-year update on a previously reported series. Sleep Med. Aug 2013;14(8):744–8. doi:10.1016/j.sleep.2012.10.009
Alonso A, Huang X, Mosley TH, Heiss G, Chen H. Heart rate variability and the risk of Parkinson's disease: The Atherosclerosis Risk in Communities (ARIC) Study. Ann Neurol. Feb 23 2015;doi:10.1002/ana.24393
Ahmadi Rastegar D, Ho N, Halliday GM, Dzamko N. Parkinson’s progression prediction using machine learning and serum cytokines. npj Parkinson's Disease. 2019/07/25 2019;5(1):14. doi:10.1038/s41531-019-0086-4
Akbilgic O, Kamaleswaran R, Mohammed A, et al. Electrocardiographic changes predate Parkinson's disease onset. Sci Rep. Jul 9 2020;10(1):11319. doi:10.1038/s41598-020-68241-6
Yuan W, Beaulieu-Jones B, Krolewski R, et al. Accelerating diagnosis of Parkinson’s disease through risk prediction. BMC Neurology. 2021/05/18 2021;21(1):201. doi:10.1186/s12883-021-02226-4
Karabayir I, Butler L, Goldman SM, et al. Predicting Parkinson's Disease and Its Pathology via Simple Clinical Variables. J Parkinsons Dis. Sep 29 2021;doi:10.3233/JPD-212876
Miyoshi F, Ogawa T, Kitao SI, et al. Evaluation of Parkinson disease and Alzheimer disease with the use of neuromelanin MR imaging and (123)I-metaiodobenzylguanidine scintigraphy. AJNR Am J Neuroradiol. Nov-Dec 2013;34(11):2113–8. doi:10.3174/ajnr.A3567
Alonso A, Huang X, Mosley TH, Heiss G, Chen H. Heart rate variability and the risk of Parkinson disease: The Atherosclerosis Risk in Communities study. Ann Neurol. May 2015;77(5):877–83. doi:10.1002/ana.24393
Chahine LM, Stern MB. Diagnostic markers for Parkinson's disease. Curr Opin Neurol. Aug 2011;24(4):309–17. doi:10.1097/WCO.0b013e3283461723
Akbilgic O, Butler L, Karabayir I, et al. ECG-AI: electrocardiographic artificial intelligence model for prediction of heart failure. European Heart Journal - Digital Health. 2021;doi:10.1093/ehjdh/ztab080
Akbilgic O, Butler L, Karabayir I, et al. Artificial Intelligence Applied to ECG Improves Heart Failure Prediction Accuracy. JACC: Journal of American College of Cardiology. 2021;77(18_Supplement_1):1.
He K, Zhang X, Ren S, Sun J. Deep residual learning for image recognition. 2016:770–778.
Maas AL, Hannun AY, Ng AY. Rectifier nonlinearities improve neural network acoustic models. Citeseer; 2013:3.
Srivastava N, Hinton G, Krizhevsky A, Sutskever I, Salakhutdinov R. Dropout: a simple way to prevent neural networks from overfitting. The journal of machine learning research. 2014;15(1):1929–1958.
Kingma DP, Ba J. Adam: A method for stochastic optimization. arXiv preprint arXiv:14126980. 2014;
Ke G, Meng Q, Finley T, et al. LightGBM: a highly efficient gradient boosting decision tree. presented at: Proceedings of the 31st International Conference on Neural Information Processing Systems; 2017; Long Beach, California, USA.

No competing interests reported.

SupplementaryMaterialScientificReportsPDAIKarabayir.docx

Download PDF

Journal Publication

published 29 Jul, 2023

Read the published version in Scientific Reports →

Editorial decision: Major revision
10 Feb, 2023
Reviews received at journal
28 Jan, 2023
Reviewers agreed at journal
25 Jan, 2023
Reviewers agreed at journal
22 Jan, 2023
Reviewers invited by journal
13 Jan, 2023
Editor assigned by journal
18 Nov, 2022
Editor invited by journal
18 Nov, 2022
Submission checks completed at journal
18 Nov, 2022
First submitted to journal
15 Nov, 2022

You are reading this latest preprint version

Externally Validated Deep Learning Model to Identify Prodromal Parkinson’s Disease from Electrocardiogram

Status:

Journal Publication

Version 1

Abstract

Figures

Introduction

Results

Discussion

Methods

Declarations

References

Additional Declarations

Supplementary Files

Status:

Journal Publication

Version 1