A machine learning-derived electrocardiographic algorithm for the detection of cardiac amyloidosis

Background

The diagnosis of cardiac amyloidosis (CA) requires advanced imaging techniques. Typical surface ECG patterns have been described, but their diagnostic value is limited.

Purpose

The aim of this study was to perform a comprehensive electrophysiological characterization in CA patients and to develop a robust, easy-to-use diagnostic tool.

Methods

First, we applied electrocardiographic imaging (ECGI) to generate detailed electroanatomical maps in CA patients and controls. Then, a machine learning approach was used to generate a surface ECG-based diagnostic algorithm from the complex dataset.

Results

Areas of low voltage were localized in the basal inferior regions of both ventricles and the remaining right ventricular segments in CA. The earliest epicardial breakthrough of myocardial activation was visualized in the right ventricle. Potential maps showed an accelerated and diffuse propagation pattern. We correlated the results from ECGI with 12-lead ECG recordings. Ventricular activation correlated best with R-peak timing in leads V1 to V3. Epicardial voltage showed a strong positive correlation with R-peak amplitude in inferior leads II, III, aVF. Ten blinded cardiologists were then asked to identify CA patients by analyzing 12-lead ECGs before and after training for the defined ECG patterns. Training resulted in significant improvements in the detection rate of CA with an AUC of 0.69 before and 0.97 after training (Figure).

Conclusion

Using a machine learning approach, a robust ECG-based tool was developed to detect CA from detailed electroanatomical mapping of CA patients. The developed tool proved to be a simple and reliable diagnostic tool to suspect CA without the aid of advanced imaging modalities.

Funding Acknowledgement

Type of funding sources: None.

P2726Extremely boosted prediction of cardiac amyloidosis by routine laboratory paramaters

Background/Introduction

Cardiac amyloidosis (CA) is a rare and complex condition with poor prognosis. Novel therapies have been shown to improve outcome, however, most of the affected individuals remain undiagnosed, mainly due to a lack in awareness among clinicians. One approach to overcome this issue is to use automated diagnostic algorithms that act based on routinely available laboratory results.

Purpose

We tested the performance of flexible machine learning and traditional statistical prediction models for non-invasive CA diagnosis based on routinely collected laboratory parameters. Since laboratory routines vary between hospitals or other health care providers, special attention has been taken to adaptive and dynamic parameter selection, and to dealing with the frequent occurrence of missing values.

Methods

Our cohort consisted of 376 clinically accepted patients with various types of heart failure. Of these, 69 were diagnosed with CA via endomyocardial biopsy (positives), and 307 had unrelated cardiac disorders (negatives). A total of 63 routine laboratory parameters were collected from these patients, with a high incidence of missing values (on average 60% of patients for each parameter). We tested the performance of two prediction models: logistic regression, and extreme gradient boosting with regression trees. To deal with missing values we adopted two strategies: a) finding an optimal overlap of parameters and deleting all patients with missing values (reduction of parameters and samples), and b) retaining all features and imputing missing values with parameter-wise means. To fairly assess the performance of prediction models we employed a 10-fold cross validation (stratified to preserve sample class ratio). Finally, area under curve for receiver-operator characteristic (ROC AUC) was used as our final performance measure.

Results

A complex machine learning model based on forests of regression trees proved to be the most performant (ROC AUC 0.94±4%) and robust to missing values. The best regression model was obtained with the 25 most frequent variables and patient deletion in case of missing values (ROC AUC 0.82±0.8%). While progressive inclusion of predictor variables worsened the performance of the logistic regression, it increased that of the machine learning approach.

Conclusions

Extreme gradient boosting of regression trees by routine laboratory parameters achieved staggering accuracy results for the automated diagnosis of CA. Our data suggest that implementations of such algorithms as independent interpreters of routine laboratory results may help to establish or suggest the diagnosis of CA in patients with heart failure symptoms, even in the absence of specialized experts.

Abstract P5-16-06: Prepectoral versus retropectoral implant-based breast reconstruction - The surgical and radiotherapeutical perspective

Background: Subpectoral implant positioning has been the standard of care in breast reconstruction despite involving disadvantages owing to the detachment of the pectoralis major muscle such as disruption of the muscle function, animation deformities and prolonged postoperative pain. Refined ablative techniques as well as dermal matrices and synthetic mesh products have led to the reintroduction of subcutaneous implant-based breast reconstruction possibly avoiding the negative sequelae of pectoralis disinsertion.

Objective: The primary objective of this study was to compare procedure-related complication rates following prepectoral versus retropectoral implant-based breast reconstruction. Furthermore the effect of the implant position on the quality of post-mastectomy radiation therapy (PMRT) was analysed.

Methods: All patients who underwent an implant-based breast reconstruction after mastectomy at the Department of Obstetrics and Gynecology of the University Clinic of Vienna within the years 1.1.2013 to 31.12.2017 were included in the study. A retrospective chart review of the patients was conducted assessing parameters regarding the mastectomy, the reconstruction, complications following the reconstructive procedure, patient-associated risk factors and radiation treatment plans. Complication rates were analysed one week, one month and one year after the implant-based reconstructive operation.

Results: In total 57 patients (94 breasts) were reconstructed following a prepectoral implant-placement approach, 95 patients (149 breasts) were reconstructed with implants in a retropectoral position. The two patient cohorts did not differ significantly in the occurrence of complications including the following dehiscence, infection, seroma, secondary bleeding, necrosis, fistula, capsular contracture and rippling. No significant differences regarding reinterventions and reoperations including seroma drainage, secondary suture and reoperation following secondary hemorrhage and necrosis could be detected between the two study populations. The two surgical procedures were not associated with a different rate of implant loss.

12 (2 in the cohort of patients with prepectorally placed implants and 10 in the subgroup of patients with subpectorally positioned implants) out of 152 patients needed PMRT for oncological safety. Prepectoral versus retropectoral implant positioning did not affect breast Dmean or D90, heart Dmax or V5 or lung V20 across treatment plans.

Conclusion: The study demonstrated no inferior outcome regarding the occurrence of complications, reinterventions, reoperations and implant loss of prepectoral implant-based breast reconstruction compared to retropectoral implant positioning. Therefore, subcutaneous implant placement permits reconstruction of the breast with comparable procedure-related complication rates while avoiding disadvantages associated with the detachment of the pectoral muscle.

Regarding the radiation perspective both prepectoral and retropectoral implant positioning allow for optimal coverage of the chest wall with acceptable doses to the heart and lung.

Citation Format: Deutschmann C, Gschwantler-Kaulich D, Dorffner G, Singer C, Leser C, Kauer-Dorner D. Prepectoral versus retropectoral implant-based breast reconstruction - The surgical and radiotherapeutical perspective [abstract]. In: Proceedings of the 2018 San Antonio Breast Cancer Symposium; 2018 Dec 4-8; San Antonio, TX. Philadelphia (PA): AACR; Cancer Res 2019;79(4 Suppl):Abstract nr P5-16-06.

Orexin receptor antagonism, a new sleep-enabling paradigm: a proof-of-concept clinical trial

The orexin system is a key regulator of sleep and wakefulness. In a multicenter, double-blind, randomized, placebo-controlled, two-way crossover study, 161 primary insomnia patients received either the dual orexin receptor antagonist almorexant, at 400, 200, 100, or 50 mg in consecutive stages, or placebo on treatment nights at 1-week intervals. The primary end point was sleep efficiency (SE) measured by polysomnography; secondary end points were objective latency to persistent sleep (LPS), wake after sleep onset (WASO), safety, and tolerability. Dose-dependent almorexant effects were observed on SE, LPS, and WASO. SE improved significantly after almorexant 400 mg vs. placebo (mean treatment effect 14.4%; P < 0.001). LPS (–18 min (P = 0.02)) and WASO (–54 min (P < 0.001)) decreased significantly at 400 mg vs. placebo. Adverse-event incidence was dose-related. Almorexant consistently and dose-dependently improved sleep variables. The orexin system may offer a new treatment approach for primary insomnia.

Risk-neutral density extraction from option prices: improved pricing with mixture density networks

One of the central goals in finance is to find better models for pricing and hedging financial derivatives such as call and put options. We present a new semi-nonparametric approach to risk-neutral density extraction from option prices, which is based on an extension of the concept of mixture density networks. The central idea is to model the shape of the risk-neutral density in a flexible, nonlinear way as a function of the time horizon. Thereby, stylized facts such as negative skewness and excess kurtosis are captured. The approach is applied to a very large set of intraday options data on the FTSE 100 recorded at LIFFE. It is shown to yield significantly better results in terms of out-of-sample pricing accuracy in comparison to the basic and an extended Black-Scholes model. It is also significantly better than a more elaborate GARCH option pricing model which includes a time-dependent volatility process. From the perspective of risk management, the extracted risk-neutral densities provide valuable information for value-at-risk estimations.

Application of artificial neural networks for detection of abnormal fetal heart rate pattern: a comparison with conventional algorithms

Cardiotocography signals were sampled during labour in 53 patients. A recurrent artificial neural network with hidden layer feedback was trained and performance was compared with that of several conventional systems. Correct and false positive rates of all systems tested were calculated. To ensure that the performance of neural networks was not just caused by using different cut-off levels, the threshold used for conventional methods were also adapted and optimised. The correct positives rate of neural networks was between 0.72 and 0.9, and the false positive rate between 0.2 and 0.4. Before optimising, conventional algorithms produced a very low correct positive (0.02-0.5) and a low false positive rate (0.0-0.08). After adjusting the parameters, the tested neural networks still performed better than optimised conventional systems.

Technische Machbarkeit der Implantation eines Monorail-Stent-Systems in die Nierenarterien ohne vorherige Dilatation

PURPOSE

To evaluate the technical feasibility of the implantation of the monorail RX Herculink system into the renal arteries without pre-dilatation.

MATERIALS AND METHODS

Forty-two patients (mean age 71 years) from four centers with a total of 44 renal artery stenoses underwent implantation of the RX Herculink stent. The mean grade of the stenosis was 83.8 %, the mean length 7.5 mm. The stenoses were ostial in 38 cases and in immediate proximity to the ostium in 6 cases. The mean follow-up-period was 57 weeks (24 - 176 weeks).

RESULTS

In 42 cases, the implantation was successful without pre-dilatation. In 2 cases, pre-dilatation was carried out. In none of the cases, detachment of the stent from the balloon was observed. In one stenosis with a length of 17 mm, implantation of two stents was performed. In 9 cases, post-dilatation with a larger balloon or higher balloon pressure was necessary. Residual stenoses exceeding 30 % were not observed. Two patients developed local bleeding at the puncture site. During the follow-up, restenoses were observed in 5 stents after 26 to 126 weeks, which necessitated a second intervention in 3 cases (PTA in 2 cases, re-stenting in 1 case). The primary patency rate after 6 and 12 months was 0.92 +/- 0.056 according to Kaplan-Meier, the secondary patency rate after 6 and 12 months was 1.0 +/- 0.0.

CONCLUSION

Implantation of the RX Herculink stent system into the renal arteries without pre-dilatation is technically feasible and safe. Even without pre-dilatation, the stent-system can be advanced through the stenosis without detachment. The complication rate is low. Our clinical results are comparable to previous studies.

Interrater reliability between scorers from eight European sleep laboratories in subjects with different sleep disorders

Interrater variability of sleep stage scorings is a well-known phenomenon. The SIESTA project offered the opportunity to analyse interrater reliability (IRR) between experienced scorers from eight European sleep laboratories within a large sample of patients with different (sleep) disorders: depression, general anxiety disorder with and without non-organic insomnia, Parkinson's disease, period limb movements in sleep and sleep apnoea. The results were based on 196 recordings from 98 patients (73 males: 52.3 +/- 12.1 years and 25 females: 49.5 +/- 11.9 years) for which two independent expert scorings from two different laboratories were available. Cohen's kappa was used to evaluate the IRR on the basis of epochs and intraclass correlation was used to analyse the agreement on quantitative sleep parameters. The overall level of agreement when five different stages were distinguished was kappa = 0.6816 (76.8%), which in terms of kappa reflects a 'substantial' agreement (Landis and Koch, 1977). For different groups of patients kappa values varied from 0.6138 (Parkinson's disease) to 0.8176 (generalized anxiety disorder). With regard to (sleep) stages, the IRR was highest for rapid eye movement (REM), followed by Wake, slow-wave sleep (SWS), non-rapid eye movement 2 (NREM2) and NREM1. The results of regression analysis showed that age and sex only had a statistically significant effect on kappa when the (sleep) stages are considered separately. For NREM2 and SWS a statistically significant decrease of IRR with age has been observed and the IRR for SWS was lower for males than for females. These variations of IRR most probably reflect changes of the sleep electroencephalography (EEG) with age and gender.

Unsupervised continuous sleep analysis

One aim of the EU-funded project SIESTA was to develop a new way of describing the human sleep-wake continuum with high temporal resolution, and independent of subjective rules, to serve as an alternative to traditional sleep scoring. Here, we report new findings obtained with a fully automatic, probabilistic sleep-analyzer using Hidden Markov Models (HMMs) based on data from a single electroencephalogram (EEG) channel. HMMs allow the analysis of non-stationary time series by modeling both the probability density functions of locally stationary data and the transition probabilities between these stable states. In the context of sleep analysis, the locally stable states can be thought of as sleep stages. The sleep-wake continuum was modeled as a mixture of three different processes by defining a three-state Gaussian Observation HMM (GOHMM). No class information from human scorers was used to train the model. The probabilities of being in any of the three states at each point in time roughly indicate the amount of wakefulness, deep sleep and rapid-eye-movement (REM) sleep with a one-second time resolution. Although it was not the aim to replicate the traditional Rechtschaffen and Kales (R&K) scoring, pseudo R&K hypnograms were constructed from the probability plots in order to compare the analyzer results with classical sleep stages by human experts. We expected that the analyzer would be able to classify data correctly from the "cornerstones" of human sleep (wakefulness, deep sleep, and REM sleep). Contrary to our previous efforts, we trained the HMMs on data from two different sleep laboratories separately, instead of generalizing data from diverse laboratories. While these stages could be detected with an accuracy of around 80% at the sleep laboratory for which we already had achieved the best results, there was no improvement from previous results by the training of a separate model in the other laboratory. This finding indicates clear laboratory effects in the signal characteristics, probably due to differences in hardware and filter settings. The presented approach, going beyond a mere replication of the traditional R&K standard, offers a continuous description of human sleep which is based on probabilistic principles. It provides a second-by-second quantification of the sleep-wake continuum and captures, although being entirely data-driven instead of rule-based, the three main processes in human sleep: wakefulness, deep sleep and REM sleep.

Quality control of polysomnographic sleep data by histogram and entropy analysis

OBJECTIVE AND METHODS

Sixteen polysomnographic recordings from 8 European sleep laboratories were analyzed. The histogram analysis was used to introduce quality control of all-night EEG recordings.

RESULTS

It was found that the header information does not always provide the real saturation values of the recording equipment. The entropy measure was used for the quantitative analysis of the dynamic range of routinely used polysomnographic recorders. It was found that the recording equipment provides EEG data with entropy in the range of 8-11 bits.

CONCLUSION

In the all-night sleep EEG were observed non-linearities. It is recommended that the equipment provide the saturation values in order to apply automated overflow detection.

Artifact processing in computerized analysis of sleep EEG - a review

Quantitative analysis of sleep EEG data can provide valuable additional information in sleep research. However, analysis of data contaminated by artifacts can lead to spurious results. Thus, the first step in realizing an automatic sleep analysis system is the implementation of a reliable and valid artifact processing strategy. This strategy should include: (1) high-quality recording techniques in order to minimize the occurrence of avoidable artifacts (e.g. technical artifacts); (2) artifact minimization procedures in order to minimize the loss of data by estimating the contribution of different artifacts in the EEG recordings, thus allowing the calculation of the 'corrected' EEG (e.g. ocular and ECG interference), and finally (3) artifact identification procedures in order to define epochs contaminated by remaining artifacts (e.g. movement and muscle artifacts). Therefore, after a short description of the types of artifacts in the sleep EEG and some typical examples obtained in different sleep stages, artifact minimization and identification procedures will be reviewed.

Neural networks for recognizing patterns in cardiotocograms

The cardiotocogram (CTG) is commonly used for routine fetal monitoring in the delivery room. A major problem is that the interpretation of the CTG trace requires experienced specialists. In order to avoid long gaps between the detection of a suspicious pattern and the intervention, the CTG has to be checked in short intervals. An automated monitoring system at the obstetric site can reduce such delays. Therefore, an alarm system immediately reporting suspicious events has been built. The focus of our study was put on the question whether AI techniques such as neural networks are suited to the task of recognizing patterns in the CTG trace. In a comparative study, their performance was evaluated against that of conventional methods. The neural networks turned out to provide significantly better results than the tested conventional methods.

An artificial neural network-based noninvasive detector for suction and left atrium pressure in the control of rotary blood pumps: an in vitro study

Rotary blood pumps are used in clinical applications to assist circulation via pumping blood from the left atrium to the aorta. Negative inflow pressures at high flow rates can cause suction of the cannula in the left atrium with deleterious effects on the atrial wall, the blood, and the lung. Therefore, stable and reliable detection of suction and the prediction of the left atrium pressure (LAP) would be of major interest for the control of these pumps. This work reports about an in vitro study of such a detector based on artificial neural networks (ANN). In the first project phase, an ANN was used to estimate the LAP based on pump speed, pump flow, and aortic pressure, obtained from a mock circulation. The inputs for the ANN were 11 characteristic values computed from these three parameters. In the second phase, another ANN was trained to classify various system states, such as suction, danger of suction (a state close to actual suction), and no suction. The first ANN was able to estimate the LAP with an accuracy of +/- 1.8 mm Hg. The discrimination of suction versus the other two states could be performed with a sensitivity and specificity of about 95% while the more interesting task of distinguishing danger of suction from no suction reached a sensitivity and specificity of about 65% (leaving 25% of each class unclassified and 10% of each class incorrectly classified).(ABSTRACT TRUNCATED AT 250 WORDS)

Automated interpretation of planar thallium-201-dipyridamole stress-redistribution scintigrams using artificial neural networks

To develop an automated image interpretation system of planar cardiac 201Tl dipyridamole stress/redistribution scintigrams, the authors used artificial neural networks that associate patterns of segmental myocardial thallium uptake with a diagnostic assessment about the presence, severity and localization of significant coronary artery disease.

METHODS

Artificial neural networks were trained and evaluated using the results from segmental thallium analysis and either expert readings in 159 cases or coronary angiography in a subgroup of 81 patients.

RESULTS

Based on receiver operating characteristics analysis, the sensitivity for the detection of significant coronary artery disease at a specificity of 90% was 51% compared with angiography and 72% compared with the human expert. For severity and localization of disease, two vascular territories assigned to the vascular bed of the left anterior descending (LAD) artery and to the territory subtended by the left circumflex artery and the right coronary artery together (CX/RCA) were included in the analysis.

CONCLUSION

Artificial neural networks may be useful to develop automated computer-based image interpretation systems of 201Tl perfusion scintigrams. However, utilization of large training datasets appears to be a prerequisite to achieve adequate diagnostic performance.

On using feedforward neural networks for clinical diagnostic tasks

In this paper we present an extensive comparison between several feedforward neural network types in the context of a clinical diagnostic task, namely the detection of coronary artery disease (CAD) using planar thallium-201 dipyridamole stress-redistribution scintigrams. We introduce results from well-known (e.g. multilayer perceptrons or MLPs, and radial basis function networks or RBFNs) as well as novel neural network techniques (e.g. conic section function networks) which demonstrate promising new routes for future applications of neural networks in medicine, and elsewhere. In particular we show that initializations of MLPs and conic section function networks--which can learn to behave more like an MLP or more like an RBFN--can lead to much improved results in rather difficult diagnostic tasks.