Classification of organised atrial arrythmias using explainable artificial intelligence

Funding Acknowledgements

Type of funding sources: Public grant(s) – National budget only. Main funding source(s): BHF

NIHR

Background

Accurately determining atrial arrhythmia mechanisms from a 12-lead ECG can be challenging. Given the high success rate of cavotricuspid isthmus (CTI) ablation, accurate identification of CTI-dependent typical atrial flutter (AFL) is important for treatment decisions and procedure planning. Machine learning, with convolutional neural networks (CNNs) in particular, has been used to classify arrhythmias using the 12-lead ECG with great accuracy. However, most studies use human interpretation of the ECG as the ground truth to label the arrhythmia ECGs. Therefore, these neural networks can only ever be as good as expert human interpretation. We hypothesised a convolutional neural network could be trained to match or even exceed expert human performance in classifying CTI-dependent AFL vs. non-CTI dependent atrial tachycardia (AT), when using findings from the invasive electrophysiology (EP) study as the gold standard.

Methods

Figure 1 summarises the study methodology. We trained a CNN on data from 231 patients undergoing EP studies for atrial tachyarrhythmia. A total of 13500 5-second 12-lead ECG segments were used for training. Each case was labelled CTI-dependent AFL or non-CTI dependent AT based on the findings of the EP study. The model performance was evaluated against a test set of 57 patients. A survey of electrophysiologists and cardiologists in Europe was undertaken on the same 57 ECGs.

Results

The model had an accuracy of 86% (95% CI 0.77-0.95). The F1 score was 0.87.The AT/AFL network correctly identified AT 82% and AFL 90% of the time.

A saliency map can be used to help understand why a CNN predicted a particular outcome. This is achieved by mapping the outcome back to key areas of the input that most influenced the network in producing the classification result. Figure 2 presents the saliency mappings of an example 12-lead ECG for each class of AFL and AT. The network used the expected sections of the ECGs for diagnoses; these were the P-wave segments and not the QRS or other unexpected segments.

There were twelve respondents in the clinician survey. These respondents included nine electrophysiologists. The median accuracy was 78% (range 70-86%). The electrophysiologists had a median accuracy of 79%, (range 70-84%). Humans were more likely to incorrectly diagnose AFL as AT (on average incorrect diagnoses: 9 AFL, 1 AT). In comparison, the neural network most often incorrectly diagnosed AT as AFL (incorrect diagnoses: 5 AT, 3 AFL).

Conclusion

We describe the first neural network trained to differentiate CTI-dependent AFL from other atrial tachycardias. We found that our model surpassed expert human performance. Automated artificial intelligence enhanced ECG analysis could help guide treatment decisions and plan ablation procedures for patients with organised atrial arrhythmias.

Diagnostic accuracy of placental growth factor and ultrasound parameters to predict the small-for-gestational-age infant in women presenting with reduced symphysis-fundus height

OBJECTIVES

To assess the diagnostic accuracy of placental growth factor (PlGF) and ultrasound parameters to predict delivery of a small-for-gestational-age (SGA) infant in women presenting with reduced symphysis-fundus height (SFH).

METHODS

This was a multicenter prospective observational study recruiting 601 women with a singleton pregnancy and reduced SFH between 24 and 37 weeks' gestation across 11 sites in the UK and Canada. Plasma PlGF concentration < 5(th) centile, estimated fetal weight (EFW) < 10(th) centile, umbilical artery Doppler pulsatility index > 95(th) centile and oligohydramnios (amniotic fluid index < 5 cm) were compared as predictors for a SGA infant < 3(rd) customized birth-weight centile and adverse perinatal outcome. Test performance statistics were calculated for all parameters in isolation and in combination.

RESULTS

Of the 601 women recruited, 592 were analyzed. For predicting delivery of SGA < 3(rd) centile (n = 78), EFW < 10(th) centile had 58% sensitivity (95% CI, 46-69%) and 93% negative predictive value (NPV) (95% CI, 90-95%), PlGF had 37% sensitivity (95% CI, 27-49%) and 90% NPV (95% CI, 87-93%); in combination, PlGF and EFW < 10(th) centile had 69% sensitivity (95% CI, 55-81%) and 93% NPV (95% CI, 89-96%). The equivalent receiver-operating characteristics (ROC) curve areas were 0.79 (95% CI, 0.74-0.84) for EFW < 10(th) centile, 0.70 (95% CI, 0.63-0.77) for low PlGF and 0.82 (95% CI, 0.77-0.86) in combination.

CONCLUSIONS

For women presenting with reduced SFH, ultrasound parameters had modest test performance for predicting delivery of SGA < 3(rd) centile. PlGF performed no better than EFW < 10(th) centile in determining delivery of a SGA infant.

Low pregnancy associated plasma protein-A in the 1st trimester: is it a predictor of poor perinatal outcome?

The objective of the study was to examine the predictive value of pregnancy associated plasma protein-A (PAPP-A) as a marker of poor pregnancy outcome. Databases at the University Hospital Lewisham, were used retrospectively to identify singleton pregnancies, which underwent 1st trimester combined screening between July 2008 and April 2010 and were found to have PAPP-A levels ≤ 0.4 MoM. The perinatal courses of these pregnancies (n = 315) were evaluated for signs of adverse perinatal outcome and compared with a matched control group of pregnancies (n = 330) with normal PAPP-A levels. Results showed that women with low serum PAPP-A were at increased risk of adverse pregnancy outcome compared with the control group (OR 2.4, p = 000.1). They were also more likely to suffer fetal loss (OR 6.2, p = 0.001) in the form of miscarriage (OR 2.7, p = 0.110) and stillbirth (OR 2.4, p = 0.001). It was concluded that serum PAPP-A is a marker for poor pregnancy outcome and women with low serum PAPP-A levels would benefit from increased monitoring of their pregnancies.

Abstract PD09-01: BRCA1 inactivation induces NF-κB in human breast cancer cells and in murine and human mammary glands

Understanding the biological mechanisms underlying the initiation and progression of breast cancer it is an important step for its prevention and treatment. In 2011 in the United States, approximately 230,000 women were diagnosed with breast cancer and 40,000 died. Individuals with mutations in breast cancer-associated gene 1 (BRCA1) have a lifetime risk of developing breast cancer up to 85%.

It is well known that BRCA1 participates in DNA damage repair and cell cycle checkpoint control, serving as a tumor suppressor gene to maintain the global genomic stability. However, BRCA1 has also been shown to play a key role in maturation of mammary stem/progenitor cells, which are the targets for carcinogenesis in individuals who have undergone loss of heterozygosity (LOH) for BRCA1. Recently, it has also been shown that NF-κB activity is increased in both mammary carcinoma cell lines and primary human breast cancer tissue. Indeed, in a previous study it has been demonstrated that NF-κB inducible kinase (NIK), p100/p52 and RelB (all components of the alternative NF-κB pathways) were increased in BRCA1-mutated tumors.

Here we show that BRCA1-loss or -mutation is responsible for activation of the alternative NF-κB pathway evidenced by NIK and IκB kinase-α (IKKα) phosphorylation, processing of p100 to p52 and p52/RelB nuclear localization. Moreover, increased p52 was also observed after BRCA1 inhibition. A BRCA1-mutated human breast cancer cell line (HCC1937) was also used to understand the role played by NIK in NF-κB alternative pathway activation. Indeed, NIK inhibition in HCC1937 cell line resulted in a decrease in p52 formation. Moreover, a decrease in NIK mRNA level was also observed when wild-type BRCA1 was reconstituted in HCC1937 cells. BRCA1 inactivation in MCF-7 cells also induced NIK phosphorylation and nuclear localization of RelB and p52. Overall, these data show that inactivation of BRCA1 increases NIK mRNA level, associated with induction of the NF-κB alternative pathway.

Stem/progenitors cells sorted using the CD24/CD49f immunophenotype derived from BRCA1 knockout mouse mammary glands showed alternative NF-κB pathway activation. Inhibition of IKKα/β using BMS-345541 completely blocked mammary colony formation in a Matrigel assay. Moreover, increased p52 formation was found in mammary stem/progenitor cells and mammary gland paraffin sections obtained from BRCA1 knockout mice. Remarkably, RelB and p100/p52 were highly expressed in 20–50% of the lobular structures in histologically normal breast tissue obtained from human BRCA1 mutation carriers while no staining was evident in normal tissue from non-carrier mastectomy samples.

Our data show that BRCA1 inactivation induces alternative NF-κB activation which ultimately promotes the expansion of the mammary progenitor population. These novel findings provide a new basis for functional classification of BRCA1 mutations and a potential method for predicting breast cancer in BRCA1 mutation carriers. Lastly our results suggest that targeting the alternative NF-κB pathway could be of benefit in the prevention of BRCA1-associated breast cancer by limiting progenitor cell expansion.

Citation Information: Cancer Res 2012;72(24 Suppl):Abstract nr PD09-01.

Antenatal detection of arteriovenous anastomoses in monochorionic twin pregnancy

OBJECTIVES

To revalidate the detection technique for arteriovenous anastomoses in an unselected group of monochorionic twins, and to make recommendations about its applicability for more widespread use.

METHODS

Women with monochorionic diamniotic (MCDA) twins were recruited and underwent placental mapping by color Doppler ultrasound. Images of placental maps showing the location and type of anastomoses were saved as digital video clips. After delivery, dye injection study of all the placentas was performed to delineate the site and type of anastomoses. A digital photograph of each injection study was taken and saved. The antenatal ultrasound images and postnatal dye injection studies were compared.

RESULTS

18 sets MCDA twins were evaluated. In 3 cases there was evidence of twin-to-twin transfusion syndrome. Dye injection of 18 placentas revealed 21 arterio-arterial anastomoses (AAA), 21 arteriovenous anastomoses (AVA) and 4 veno-venous anastomoses (VVA). Of these, 10 (48%) AAAs and 5 (24%) AVAs were detected antenatally by color Doppler. In all cases, where an AVA was detected, the placenta was located anteriorly.

CONCLUSION

Antenatal detection of AVA was feasible when the placenta was located anteriorly, but proved difficult in posteriorly situated placentas.

Azathioprine and breastfeeding: is it safe?

Traditionally, women receiving azathioprine have been discouraged from breastfeeding because of theoretical potential risks of neonatal bone marrow suppression, susceptibility to infection, and pancreatitis. The aims of this study were to measure the concentration of 6-mercaptopurine (6-MP) in breast milk of mothers receiving azathioprine and in the blood of their babies and to investigate any immunosuppressive effects on the babies. Women receiving azathioprine, who after appropriate counselling wished to breastfeed their babies, were approached for inclusion in the study. Breast milk samples were obtained from recruited women, and 6-MP levels were measured in each breast milk sample. Haemoglobin level, white cell and platelet counts, and 6-MP and 6-thioguanine nucleotides (6-TGN) levels were measured in the respective neonatal blood samples. Clinical signs of immunosuppression in the neonates were noted. Thirty-one breast milk samples were collected from ten women. Low concentrations of 6-MP (1.2 and 7.6 nanograms/ml, compared with therapeutic immunosuppressant level of 50 nanograms/ml in serum) were detected in two breast milk samples obtained from one woman. 6-MP was not detected in any of the other 29 samples. 6-MP and 6-TGN were undetectable in the neonatal blood. There were no clinical or haematological signs of immunosuppression in any of the ten neonates. We conclude that breastfeeding should not be withheld in infants of mothers receiving azathioprine.

Short communication: Vaginal delivery can be considered in monochorionic diamniotic twins

The outcomes of 60 sets of monochorionic diamniotic (MCDA) twins were compared with 218 sets of dichorionic diamniotic (DCDA) twins. The caesarean section rates for MCDA were similar to those for DCDA twins (56.6 versus 53.6%, P > 0.1). Although the number of babies with 5-minute Apgar score of <7 was significantly higher for vaginally delivered MCDA twins compared with that of DCDA twins (12 versus 3.5%, P < 0.001), the umbilical artery pH of <7.2 was similar (20 versus 13%, P > 0.05). Admission to neonatal intensive care unit (NICU) and neonatal mortality were also similar in both groups. Delivery by caesarean section was associated with increased admission to the NICU and neonatal mortality for MCDA twins when compared with vaginal delivery group. From this retrospective cohort study, we can conclude that vaginal delivery for MCDA twins appeared to be a reasonable management option when similar selection criteria for vaginal delivery of DCDA twins were applied.

Intraobserver and interobserver variation in the sonographic grading of placental maturity

OBJECTIVES

The appearance of Grannum Grade III changes in the placenta at around 34-36 weeks is a predictor of adverse perinatal outcome, which may be reduced by reporting to the clinician. This has led to the suggestion that the placental grade should be noted during any third-trimester scan. There are no published data on the reproducibility of sonographic Grannum grading of the placenta; the objective of this study was to evaluate intra- and interobserver variation.

METHODS

Fifty-five placental images from normal and complicated pregnancies of several different gestational ages were collected between April and October 2001. Three fetal medicine consultants and three experienced sonographers graded the images as 0, I, II, III or ungradeable. They then regraded the same images, presented in a different order and with different codes, 4-6 weeks later. Observers were blinded to their previous grading and to each others'. Weighted kappa (kappa), with linear weights, was used to look for strength of agreement.

RESULTS

There was good agreement between the two observations of each placental image for five observers (kappa = 0.61 to 0.90), and moderate agreement for one observer (kappa = 0.56). However, the kappa-values for comparisons between the 15 pairs of observers ranged from 0.24 to 0.69 with six values below 0.41, indicating only fair agreement. This was confirmed by the overall kappa-value of 0.24 between all six observers. The agreement between the observers for Grade III placenta was poor, with an overall kappa-value of 0.09.

CONCLUSIONS

Although intraobserver agreement was generally good, interobserver agreement was only fair for all grades and poor for Grade III placenta. This may be an indication that Grannum grading is not reproducible or it may reflect a need for training in those performing grading. Such variation may limit the effectiveness of reporting Grannum grades in clinical practice.

Spanwise bifurcation in plane-symmetric sudden-expansion flows

Present computational investigation reports a steady bifurcation phenomenon for three-dimensional flows through a plane-symmetric sudden expansion. When the channel aspect ratio exceeds a critical value, the well-known step height (pitchfork) bifurcation evolves with different symmetry breaking orientations on the left and right sides of the channel and bifurcates in the spanwise direction. For the channel aspect ratio less than the critical value, the originally occurring spanwise bifurcation cannot be stably retained and evolves eventually to a step height bifurcation. Compared to step height bifurcation, the spanwise bifurcation is found to be more difficult to obtain, because the symmetric flow present on the spanwise symmetry plane is unstable in two dimensions. For completeness, an extensive analysis of the observed spanwise bifurcation, covering its transient behavior, dependence on flow Reynolds number, channel aspect ratio, and expansion ratio, is included.

Screening for trisomy 21: the significance of a positive second trimester serum screen in women screen negative after a nuchal translucency scan

Screening of pregnancies for trisomy 21 is now an accepted part of antenatal care. Measurement of fetal nuchal translucency in the first trimester and analysis of maternal serum biochemistry in the second trimester are both established methods of screening. The performance characteristics of both tests in an unselected population are well described and the choice of test offered is usually determined by local policy and resources. We present data from a screening programme offering women with a low risk result from nuchal translucency measurement a second trimester serum screen. There were eight cases of trisomy 21 in the 2683 women screened, all of which presented with a high-risk nuchal screen result. Serum screening of 1057 women who screened negative by nuchal translucency gave 46 high-risk results, all of which were, therefore, false positive for trisomy 21. Second trimester biochemistry screening following a negative nuchal translucency screen did not increase the detection of trisomy 21.