Deep Sequence Modeling for Hemorrhage Diagnosis Conference on - - PowerPoint PPT Presentation

1 Dec 8, 2018 Fabian Falck et al. – Deep Sequence Modeling for Hemorrhage Diagnosis

Deep Sequence Modeling for Hemorrhage Diagnosis

Conference on Neural Information Processing Systems (NeurIPS) 2018 Workshop Machine Learning for Health

Fabian Falck 1, Michael R. Pinsky, MD 2, Artur Dubrawski 1

1 Auton Lab, School of Computer Science, Carnegie Mellon University 2 Dep of Critical Care Medicine, School of Med, University of Pittsburgh

Partially supported by R01 GM117622

2 Dec 8, 2018 Fabian Falck et al. – Deep Sequence Modeling for Hemorrhage Diagnosis

Medical relevance and Task definition

• Medical relevance:

– Hemorrhage (internal bleeding) is the most common cause of trauma deaths and the most frequent complication of major surgery – With current medical practice, it is very difficult to identify until profound blood loss has already occurred due to interrelated reflexes of the body

• Task: Predicting hemorrhage based on physiological reactions in waveform

(250 Hz) vital sign without manual featurization

3 Dec 8, 2018 Fabian Falck et al. – Deep Sequence Modeling for Hemorrhage Diagnosis

Dataset

• Controlled, constant blood loss on

93 healthy pigs transitioning from stable to unstable [1]

• Monitoring of 11 vital signs:

– blood pressure (CVP, arterial pressure fluid filled and millar, pulmonary pressure) – oxygen saturations (SpO2, SvO2) – EKG, Plethysmograph, CCO, stroke volume variation (Vigeleo), airway pressure

• Chosen for experiments: 16 pigs bleeding at

the slowest rate (5 mL/min) à hardest task

t=30min

label pred.

Start of bleeding t=60min t=0min

[1] The study protocol was approved by the University of Pittsburgh IACUC

4 Dec 8, 2018 Fabian Falck et al. – Deep Sequence Modeling for Hemorrhage Diagnosis

Models and Experimental setup

• Models: Gated Recurrent Unit (GRU) and dilated, causal convolution
• Experimental setup:

– Training on randomly drawn time windows of length W (~20sec = 5000 dp.) – Subject-specific normalization (per vital sign), k=5-fold cross-validation

Inputs Outputs Neural network layer with activation function T ensor flow T ensor concatenation T ensor duplication Pointwise multiplication Pointwise addition Pointwise operation

x +

1-

Gated Recurrent Unit (GRU) Dilated, causal convolution

x x + x

1- f

f f f f f f

Convolutional filter

f f f f f f f f

+ + + + + + +

5 Dec 8, 2018 Fabian Falck et al. – Deep Sequence Modeling for Hemorrhage Diagnosis

Experimental results

• GRU-based model achieves best performance in small false-positive

range, while being inferior for negatives compared to a formidable baseline using manually extracted features and a random forest classifier Ours AUC (positives): 0.94 TPR@FPR=0.1%: 0.38 TNR@FNR=1%: 0.22 Baseline (RF + handcrafted features) AUC (positives): 0.97 TPR@FPR=0.1%: 0.15 TNR@FNR=1%: 0.64

6 Dec 8, 2018 Fabian Falck et al. – Deep Sequence Modeling for Hemorrhage Diagnosis

Conclusion

• Limitations:

– 1) Potential biases due to lab draws and artefacts in the data might be root cause for inferior negative performance – 2) Raw predictions are noisy

• Future work:

– 1) Bayesian update scheme for raw classifier predictions – 2) Dilated, causal CNN as feature extraction head for subsequent RNN

• … and for details, please come to my poster in session 2!