The effect of confounding data features on a deep learning algorithm to predict complete coronary occlusion in a retrospective observational setting

Aims

Deep learning (DL) has emerged in recent years as an effective technique in automated ECG analysis.

Methods and results

A retrospective, observational study was designed to assess the feasibility of detecting induced coronary artery occlusion in human subjects earlier than experienced cardiologists using a DL algorithm. A deep convolutional neural network was trained using data from the STAFF III database. The task was to classify ECG samples as showing acute coronary artery occlusion, or no occlusion. Occluded samples were recorded after 60 s of balloon occlusion of a single coronary artery. For the first iteration of the experiment, non-occluded samples were taken from ECGs recorded in a restroom prior to entering theatres. For the second iteration of the experiment, non-occluded samples were taken in the theatre prior to balloon inflation. Results were obtained using a cross-validation approach. In the first iteration of the experiment, the DL model achieved an F1 score of 0.814, which was higher than any of three reviewing cardiologists or STEMI criteria. In the second iteration of the experiment, the DL model achieved an F1 score of 0.533, which is akin to the performance of a random chance classifier.

Conclusion

The dataset was too small for the second model to achieve meaningful performance, despite the use of transfer learning. However, 'data leakage' during the first iteration of the experiment led to falsely high results. This study highlights the risk of DL models leveraging data leaks to produce spurious results.

A timely diagnosis of myocardial infarction

The diagnosis of acute myocardial infarction currently rests on the measurement of troponin, a biomarker of myocardial necrosis. Unfortunately, the current generation troponin assays detect troponin only 6-9 h after symptom onset. This can lead to a delay in diagnosis and also excessive resource utilization when triaging patients who, ultimately, have noncardiac causes of acute chest pain. For these reasons, there has been extensive research interest in biomarkers that can detect and rule out myocardial infarction early after symptom onset. These include markers of myocardial injury, such as myoglobin, heart-type fatty acid binding protein, glycogen phosphorylase BB; hemostatic markers, such as D-dimer; and finally, inflammatory markers, such as matrix metalloproteinase 9. Recently, highly sensitive troponin assays have reported an early sensitivity for myocardial infarction of greater than 95%, although at a cost of reduced specificity. The optimal strategy with which to use these novel biomarkers and highly sensitive troponins has yet to be determined, and interpretation of their results in light of thorough clinical assessment remains essential.

Biphasic energy selection for transthoracic cardioversion of atrial fibrillation. The BEST AF Trial

AIMS

To compare the efficacy and safety of an escalating energy protocol with a non-escalating energy protocol using an impedance compensated biphasic defibrillator for direct current cardioversion of atrial fibrillation (AF).

METHODS AND RESULTS

This prospective multicentre randomised trial enrolled 380 patients (248 male, mean (SD) age 67 (10) years) with AF. Patients were randomised to either an escalating energy protocol (protocol A: 100 J, 150 J, 200 J, 200 J), or a non-escalating energy protocol (protocol B: 200 J, 200 J, 200 J). Cardioversion was performed using an impedance compensated biphasic waveform. First-shock success was significantly higher for those randomised to 200 J than 100 J (71% vs 48%; p<0.01) and for patients with a body mass index (BMI) >25 kg/m(2) (75% vs 44%; p = 0.01). In patients with a normal BMI there was no significant difference in first-shock success. There was also no significant difference between subsequent shocks or overall success. The use of a non-escalating protocol (protocol B) resulted in fewer shocks but with a higher cumulative energy. There was no difference in duration of procedure, amount of sedation administered or post-shock erythema between the groups.

CONCLUSION

First-shock success was significantly higher, particularly in patients with a BMI >25 kg/m(2), when a non-escalating initial 200 J energy was selected. The overall success, duration of procedure and amount of sedation administered, however, did not differ significantly between the two protocols.

Neuropeptide variability in man

BACKGROUND

Previous studies have established short-term variability in the circulating plasma levels of cardiac peptides such as atrial natriuretic peptide (ANP) and brain natriuretic peptide (BNP). Our aim was to investigate whether such variable patterns could be observed in other vasoactive peptides.

METHODS

We measured the immunoreactivity of vasoactive intestinal polypeptide (VIP), neuropeptide Y (NPY), endothelin-1 (ET-1) and calcitonin gene-related peptide (CGRP) in peripheral venous plasma collected at 2-min intervals over a 20-min period from patients with chronic cardiac failure (CCF) and from control subjects. In a second study, blood samples were obtained at 2-min intervals from the pulmonary artery, femoral artery and antecubital vein from patients with normal cardiac function while right atrial pressure and heart rate were constant.

RESULTS

Peripheral blood VIP, NPY and ET-1 had peaks and troughs (levels > 2SD from the mean) in both patients and controls, with approximate intervals of 10 min. Levels of CGRP showed little variation. The overall levels [median (range); pmol L-1] of VIP [patients 27 (2.1-85.5); controls 9.8 (0-34)] and NPY [patients 20 (0-110); controls 12 (5-19)] were higher in patients (P < 0.05). Circulating plasma levels of ET-1 and CGRP were about the same in both groups [ET-1: patients 18 (2-84); controls 18 (0-48); CGRP: patients 4 (1-18.5), controls 5.5 (1-15); P = NS]. Levels of CGRP, VIP and ET-1 were similar in the pulmonary and femoral arteries, whereas systemic arterial levels of NPY were higher than in the pulmonary artery.

CONCLUSIONS

The data demonstrate marked variability in circulating levels of the neuropeptides studied. In addition, peaks and troughs were observed every 10-15 min from all three vascular beds. If these peptides are secreted in a pulsatile pattern, then interpretations of single measurements should be guarded. Furthermore, this study raises interesting questions about the physiology of hormone secretion in man.

A gastroesophageal electrode for electrophysiological studies

A novel gastroesophageal electrode has been developed capable of atrial and ventricular pacing. We performed electrophysiological studies using the gastroesophageal electrode (Esothoracic) and compared the results with the standard endocardial approach. The flexible polythene gastroesophageal electrode was passed into the stomach under light sedation. Five ring electrodes, now positioned in the lower esophagus were used for bipolar atrial pacing and recording. Ventricular pacing was performed using a cathodic point source on the gastroesophageal electrode tip; the indifferent electrode (anode) was a high impedance chest pad. Parameters of sinus and AV nodal function were obtained by atrial pacing. Programmed ventricular stimulation was performed using a standard protocol. These electrophysiological parameters were subsequently determined using the endocardial approach. There was close correlation between measurements of sinus and AV node function using the two approaches in 48 subjects: sinus node recovery time (SNRT) r2 = 0.70, corrected sinus node recovery time (CSNRT) r2 = 0.87, AV Wenckebach cycle length (AVWCL) r2 = 0.97. The degree of agreement between the two approaches was estimated by the mean difference delta and standard deviation of the difference sigma (SNRT delta = 40 ms, sigma = 257 ms; CSNRT sigma = 14 ms, delta = 164 ms; AVWCL sigma = 7 ms, delta = 16 ms). Programmed ventricular stimulation was performed in 15 of 48 subjects with known or suspected ventricular tachyarrhythmias. Seven had ventricular tachycardia induced using both esothoracic and endocardial programmed ventricular stimulation. One subject was noninducible using esothoracic programmed ventricular stimulation, but inducible at endocardial electrophysiological studies. Another subject was inducible at esothoracic electrophysiological studies, but noninducible using endocardial programmed ventricular stimulation. Six subjects were noninducible using both endocardial and esothoracic programmed ventricular stimulation. The gastroesophageal electrode permits reliable atrial and ventricular pacing without transvenous catheterization or fluoroscopy. Electrophysiological parameters determined using this electrode are similar to those obtained using endocardial stimulation.

A gastroesophageal electrode for atrial and ventricular pacing

Temporary transvenous cardiac pacing requires technical expertise and access to fluoroscopy. We have developed a gastroesophageal electrode capable of atrial and ventricular pacing. The flexible polythene gastroesophageal electrode is passed into the stomach under light sedation. Five ring electrodes, now positioned in the lower esophagus, are used for atrial pacing. A point source (cathode) on the distal tip of the electrode, now positioned in the gastric fundus, is used for ventricular pacing. Two configurations of atrial and ventricular pacing were compared: unipolar and bipolar. During unipolar ventricular pacing the indifferent electrode (anode) was a high impedance chest pad. For bipolar ventricular pacing the indifferent electrode was a ring electrodes placed 2 cm proximal to the tip. Unipolar atrial pacing was performed with 1 of 5 proximal ring electrodes acting as cathode ("cathodic") or as anode ("anodic") in conjunction with a chest pad. Bipolar atrial pacing was performed using combinations of 2 of 5 ring electrodes. Atrial capture was obtained in all 55 subjects attempted. When all electrode combinations were compared, atrial capture was significantly more frequent using the bipolar approach (153/210 bipolar, 65/210 unipolar; t = 7.37, P < 0.001). For unipolar atrial pacing, cathodic stimulation (from esophagus) was more successful than anodic stimulation (cathodic 62/105, anodic 20/105; t = 5.81, P < 0.001). In 43 subjects attempted unipolar ventricular pacing resulted in a higher frequency of capture than the bipolar approach (unipolar 41/43 (95.3%), bipolar 19/43 (44.2%); P < 0.001). In conclusion, atrial pacing was optimal using pairs of ring electrodes ("bipolar") while ventricular pacing was optimal using the distal electrode tip (cathode) in conjunction with a chest pad electrode ("unipolar"). This gastroesophageal electrode may be useful in the emergency management of acute bradyarrhythmias and for elective electrophysiological studies.

Ventricular pacing with a novel gastroesophageal electrode: a comparison with external pacing

Temporary endocardial pacing is a technically demanding invasive procedure requiring sterile precautions and access to fluoroscopy. External (transcutaneous) pacing requires high current for capture and is poorly tolerated in the conscious patient. An esothoracic pacing system has been developed capable of reliable ventricular capture. The flexible gastroesophageal electrode is passed into the stomach. The distal 6 cm is angled to 90 degrees with an internal pulley system, positioning the tip of the gastroesophageal electrode in the fundus of the stomach. Ventricular pacing is performed with a spherical electrode (cathode) mounted on the gastroesophageal electrode tip in conjunction with a chest pad (anode) positioned medial to the cardiac apex. Of 91 subjects in which esothoracic pacing was attempted, 86 (94.5%) demonstrated successful ventricular capture at the maximum pulse duration used (40 msec). Threshold current for ventricular capture ranged from 22.5 +/- 8.1 mA at a pulse duration of 40 msec to 29.9 +/- 8.6 mA at a pulse duration of 10 msec. Esothoracic pacing was compared with external pacing in a subgroup (n = 30) of patients. Ventricular capture with the gastroesophageal electrode was more common when compared with the external approach (27 [90%] of 30 vs 13 [43.3%] of 30, p < 0.001). In those subjects in whom ventricular capture was obtained with both methods, threshold current for capture was significantly lower with the esothoracic approach. This gastroesophageal electrode may be useful in the emergency management of acute bradyarrhythmias.

An esophageal and gastric approach to ventricular pacing

Using a unipolar esothoracic pacing system (where current passes from a point source positioned in the distal esophagus to a chest wall pad) and pulse duration of 50 msec, satisfactory 1:1 ventricular capture was obtained in 57 (86%) of 66 patients, with a mean threshold current of 27.7 mA at an optimal depth of 40.3 cm from the lower lip. When the unipolar esothoracic and bipolar transesophageal ventricular pacing systems were compared, the bipolar system was associated with a lower success rate and higher threshold current. When unipolar esothoracic pacing and gastrothoracic pacing (where current passes from a point source positioned in the stomach to a chest wall pad) were compared in 23 patients with bradyarrhythmia, ventricular capture was achieved using gastrothoracic pacing in 22 patients (96%) and esothoracic pacing in 21 (91%): gastrothoracic pacing required less current (16.0 mA +/- SD 7.2 vs 25.8 mA +/- SD 8.6). Optimal ventricular capture occurred using a unipolar gastrothoracic pacing electrode inserted to an average depth of 44.3 cm together with a high impedance chest pad (250 omega) placed in the fourth interspace at the left sternal edge, with 50-msec current pulses and a mean threshold of 16.0 mA. Thus, using a gastroesophageal electrode system, ventricular pacing can be achieved successfully, and the availability of such a system could play a major role in resuscitation of patients from severe bradyarrhythmias.

Use of the gastro-oesophageal route for the rapid establishment of ventricular pacing

A novel approach to emergency ventricular pacing has been developed using a gastro-esophageal electrode. The polythene electrode was passed into the stomach, after which the electrode tip was positioned in the gastric fundus. Ventricular pacing was performed using a cathode mounted on the electrode tip; the indifferent electrode (anode) was either a chest pad or a proximal ring electrode. Ventricular capture was easily achieved in three emergency cases of severe bradyarrhythmia.

Variable patterns of atrial natriuretic peptide secretion in man

Peripheral circulating levels of atrial natriuretic peptide may exhibit short-term variation compatible with a pulsatile pattern of secretion. We obtained samples every 2 min for 90 min from the antecubital vein of 16 patients with chronic cardiac failure and 13 controls. Overall levels were higher in the patients (median and quartiles 230 (125,325) vs. 26 (16,48) ng l-1; P < 0.001). In both groups there was considerable variability, with 10 (2-12) peaks, 9 (7-15) troughs (both defined as > 2 SD from the mean) and 16 (13-18) pulses (defined by computer) during the sampling period in controls, and a similar number in patients. We then carried out simultaneous sampling in the pulmonary artery, femoral artery and peripheral vein in eight subjects with normal cardiac function and six patients with impaired function due to valvular heart disease. The pattern of variability was preserved in all three sites in both groups, suggesting intermittent secretion rather than variable breakdown of the peptide in the lung. No changes in right atrial pressure or heart rate were observed to coincide with the variations, but levels of the peptide in the pulmonary artery correlated with right atrial pressure in patients (r = 0.87; P < 0.05). The mechanism of such periodicity and its pathophysiological importance remain unknown.

Non-invasive measurement of cardiac output by a carbon dioxide rebreathing method at rest and during exercise

Cardiac output was measured in 11 patients undergoing routine cardiac catheterization using a carbon dioxide rebreathing technique and compared with cardiac output measured by direct Fick and thermodilution. The carbon dioxide rebreathing technique gave consistently lower values for cardiac output than the other two methods (mean difference -0.73, 95% CI -0.95 to -0.51 l.min-1 with the direct Fick and -0.72, 95% CI -1.19 to -0.26 l.min-1 with thermodilution). The direct Fick and thermodilution methods gave similar results (mean difference -0.08, 95% CI -0.32 to 0.16 l.min-1). Cardiac output was also measured in 10 healthy subjects at rest and during two steady-state levels of exercise using the carbon dioxide rebreathing technique. Measurements were made in triplicate on 3 separate days. The technique gave reproducible results between replicates at rest (coefficient of variation 9.1%) and became more reproducible on exercise (coefficients of variation 5.6% and 5.4% respectively at each exercise level). There was a good correlation between cardiac output and oxygen consumption (r = 0.98). The carbon dioxide rebreathing technique is a feasible non-invasive way of measuring cardiac output. It tends to underestimate cardiac output at rest but is reproducible and becomes more so on exercise which is where it should be of most value.

Transoesophageal versus transchest DC cardioversion

Attempted cardioversion via the oesophagus (transoesophageal cardioversion) was compared with the transchest approach (transchest cardioversion) in a randomized trial of 100 consecutive patients with atrial fibrillation. For the transoesophageal group, 30, 50 and 100 J were delivered via an oesophageal electrode with subsequent 200 and 360 J transchest if required. For the transchest group, 50, 100, 200 and 360 J were delivered if required. In the transoesophageal group, 36/50 (72%) of patients cardioverted using the transoesophageal route alone, and in the transchest group, 41/50 (82%) of patients cardioverted (p = NS). First shock success was similar for the transoesophageal and transchest groups: 13/50 (26%) vs. 8/50 (16%) respectively. The mean number of shocks required to achieve successful cardioversion was identical for the transoesophageal and transchest groups (2.6). However, transoesophageal cardioversion was more successful than transchest cardioversion at energies < or = 100 J (36/50 [72%], and 17/50 [34%], p < 0.05). Median total energy for successful cardioversion was lower for patients in the transoesophageal group (180 J) than the transchest group (350 J) and mean peak current at successful cardioversion was also lower for patients in the transoesophageal group (21.7 A) than the transchest group (27.3 A) (p < 0.05). No oesophageal complications occurred. Thus, using an oesophageal electrode, cardioversion can be achieved as successfully as using the transchest route. The transoesophageal approach offers a low impedance, and consequently a low-energy pathway for cardioversion.