Machine learning-based dynamic mortality prediction after traumatic brain injury.

Our aim was to create simple and largely scalable machine learning-based algorithms that could predict mortality in a real-time fashion during intensive care after traumatic brain injury. We performed an observational multicenter study including adult TBI patients that were monitored for intracranial pressure (ICP) for at least 24 h in three ICUs. We used machine learning-based logistic regression modeling to create two algorithms (based on ICP, mean arterial pressure [MAP], cerebral perfusion pressure [CPP] and Glasgow Coma Scale [GCS]) to predict 30-day mortality. We used a stratified cross-validation technique for internal validation. Of 472 included patients, 92 patients (19%) died within 30 days. Following cross-validation, the ICP-MAP-CPP algorithm's area under the receiver operating characteristic curve (AUC) increased from 0.67 (95% confidence interval [CI] 0.60-0.74) on day 1 to 0.81 (95% CI 0.75-0.87) on day 5. The ICP-MAP-CPP-GCS algorithm's AUC increased from 0.72 (95% CI 0.64-0.78) on day 1 to 0.84 (95% CI 0.78-0.90) on day 5. Algorithm misclassification was seen among patients undergoing decompressive craniectomy. In conclusion, we present a new concept of dynamic prognostication for patients with TBI treated in the ICU. Our simple algorithms, based on only three and four main variables, discriminated between survivors and non-survivors with accuracies up to 81% and 84%. These open-sourced simple algorithms can likely be further developed, also in low and middle-income countries.

Assessment of myocardial infarct size with body surface potential mapping: validation against contrast-enhanced cardiac magnetic resonance imaging.

BACKGROUND: Assessment of myocardial infarct (MI) size is important for therapeutic and prognostic reasons. We used body surface potential mapping (BSPM) to evaluate whether single-lead electrocardiographic variables can assess MI size. METHODS: We performed BSPM with 120 leads covering the front and back chest (plus limb leads) on 57 patients at different phases of MI: acutely, during healing, and in the chronic phase. Final MI size was determined by contrast-enhanced cardiac magnetic resonance imaging (DE-CMR) and correlated with various computed depolarization- and repolarization-phase BSPM variables. We also calculated correlations between BSPM variables and enzymatic MI size (peak CK-MBm). RESULTS: BSPM variables reflecting the Q- and R wave showed strong correlations with MI size at all stages of MI. R width performed the best, showing its strongest correlation with MI size on the upper right back, there representing the width of the "reciprocal Q wave" (r = 0.64-0.71 for DE-CMR, r = 0.57-0.64 for CK-MBm, P < 0.0001). Repolarization-phase variables showed only weak correlations with MI size in the acute phase, but these correlations improved during MI healing. T-wave variables and the QRSSTT integral showed their best correlations with DE-CMR defined MI size on the precordial area, at best r = -0.57, P < 0.0001 in the chronic phase. The best performing BSPM variables could differentiate between large and small infarcts at all stages of MI. CONCLUSIONS: Computed, single-lead electrocardiographic variables can estimate the final infarct size at all stages of MI, and differentiate large infarcts from small.

Investigations of sensitivity and resolution of ECG and MCG in a realistically shaped thorax model.

Solving the inverse problem of electrocardiography (ECG) and magnetocardiography (MCG) is often referred to as cardiac source imaging. Spatial properties of ECG and MCG as imaging systems are, however, not well known. In this modelling study, we investigate the sensitivity and point-spread function (PSF) of ECG, MCG, and combined ECG+MCG as a function of source position and orientation, globally around the ventricles: signal topographies are modelled using a realistically-shaped volume conductor model, and the inverse problem is solved using a distributed source model and linear source estimation with minimal use of prior information. The results show that the sensitivity depends not only on the modality but also on the location and orientation of the source and that the sensitivity distribution is clearly reflected in the PSF. MCG can better characterize tangential anterior sources (with respect to the heart surface), while ECG excels with normally-oriented and posterior sources. Compared to either modality used alone, the sensitivity of combined ECG+MCG is less dependent on source orientation per source location, leading to better source estimates. Thus, for maximal sensitivity and optimal source estimation, the electric and magnetic measurements should be combined.

Comparison of minimum-norm estimation and beamforming in electrocardiography with acute ischemia.

In the electrocardiographic (ECG) inverse problem, the electrical activity of the heart is estimated from measured electrocardiogram. A model of thorax conductivities and a model of the cardiac generator is required for the ECG inverse problem. Limitations and errors in methods, models, and data will lead to errors in the estimates. However, in experimental applications, the use of limited or erroneous models is often inevitable due to necessary model simplifications and the difficulty of obtaining accurate 3D anatomical imaging data. In this work, we focus on two methods for solving the inverse problem of ECG in the case of acute ischemia: minimum-norm (MN) estimation and linearly constrained minimum-variance beamforming. We study how these methods perform with different sizes of ischemia and with erroneous conductivity models. The results indicate that the beamformer can localize small ischemia given an accurate model, but it cannot be used for estimating the size of ischemia. The MN estimator is tolerant to geometry errors and excels in estimating the size of ischemia, although the beamformer performs better with accurate model and small ischemia.

Predicting recovery of myocardial function by electrocardiography after acute infarction.

BACKGROUND: In acute ischemic left ventricular (LV) dysfunction, distinguishing viable myocardium is clinically important. METHODS: Body surface potential mapping (Electrocardiography [ECG] with 123 leads), was recorded in 62 patients with acute coronary syndrome (ACS). ECG variables were computed from de- and repolarization phases. LV segmental wall motion was assessed by echocardiography acutely and after 1 year. RESULTS: The number of dysfunctional segments (DFS) diminished during follow-up in 37 patients (recovery group) and remained the same or increased in 25 patients (nonrecovery group). Acutely, DFS was 5.7 ± 2.1 versus 4.4 ± 2.4 (P = 0.02), and peak CK-MBm 141 ± 157 versus 156 ± 167 µg/L (P = 0.78) in the recovery versus nonrecovery group. At follow-up, DFS was 1.9 ± 1.7 versus 6.5 ± 2.6 (P < 0.001). The best ECG variable to predict decrease in DFS depended on the region of acute LV dysfunction: The best variable in the left anterior descending region was the integral of the first QRS integral (area under the curve [AUC] 0.82, P = 0.002); in the right coronary artery region, this was the integral of the ST segment (AUC 0.98, P = 0.003); and in the left circumflex region, the area including the ST segment and the T wave (AUC 0.97, P = 0.006). CONCLUSIONS: In ACS patients, computed ECG variables predict recovery of LV function from ischemic myocardial injury, even in the presence of comparable CK-MBm release and LV dysfunction.

Fragmented QRS in prediction of cardiac deaths and heart failure hospitalizations after myocardial infarction.

BACKGROUND: Increased QRS fragmentation in visual inspection of 12-lead ECG has shown association with cardiac events in postmyocardial infarction (MI) patients. We investigated user-independent computerized intra-QRS fragmentation analysis in prediction of cardiac deaths and heart failure (HF) hospitalizations after MI. METHODS: Patients (n = 158) with recent MI and reduced left ventricular ejection fraction (LVEF) were studied. A 120-lead body surface potential mapping was performed at hospital discharge. Intra-QRS fragmentation was computed as the number of extrema (fragmentation index FI) in QRS. QRS duration (QRSd) was computed for comparison. RESULTS: During a mean follow-up of 50 months 15 patients suffered cardiac death and 23 were hospitalized for HF. Using the mean + 1 SD as cut-point both parameters were univariate predictors of both end-points. In multivariate analysis including age, gender, LVEF, previous MI, bundle branch block, atrial fibrillation, and diabetes FI was an independent predictor for cardiac deaths (HR 8.7, CI 3.0-25.6) and HF hospitalizations (HR 3.8, CI 1.6-9.3) whereas QRSd only predicted HF hospitalizations (HR 4.6, CI 2.0-10.7). In comparison to QRSd, FI showed better positive (PPA) and equal negative (NPA) predictive accuracy for both end-points, and PPA was further improved when combined to LVEF < 40%. Limiting fragmentation analysis to 12-lead ECG or a randomly selected 8-lead set instead of all 120 leads resulted in an almost similar prediction. CONCLUSIONS: Increased QRS fragmentation in post-MI patients predicts cardiac deaths and HF progression. A computer-based fragmentation analysis is a stronger predictor than QRSd.

Complex T-wave morphology in body surface potential mapping in prediction of arrhythmic events in patients with acute myocardial infarction and cardiac dysfunction.

AIMS: Heterogeneous ventricular repolarization is associated with sudden cardiac death after myocardial infarction (MI). This prospective study investigated repolarization disparity with parameters based on T-wave morphology in body surface potential mapping (BSPM) in the assessment of arrhythmia risk in patients with a recent MI and cardiac dysfunction. METHODS AND RESULTS: Patients (n = 158) had 120-lead BSPM and 12-lead electrocardiogram (ECG) registered soon after acute MI. Principal component analysis (PCA) of the T-wave and T-wave vector loop descriptors were applied to compute parameters describing T-wave morphology and its variation. The study endpoints were arrhythmic events and all-cause mortality. During a mean follow-up of 50 months, 30 patients (19%) died and 16 (10%) had an arrhythmic event. Most of the parameters differed significantly between patients with and without arrhythmic events. In univariate analysis, T-wave vector loop length (TLL) and PCA parameter PCA(3) in BSPM and TLL in ECG were significant predictors of arrhythmic events. In multivariate analysis including several clinical variables, these parameters also showed an independent prediction, with parameters in BSPM performing somewhat better. None of the parameters predicted all-cause mortality. CONCLUSION: Complex T-wave morphology in BSPM is a marker of arrhythmia propensity in patients with a recent MI and cardiac dysfunction.