Figure 9 - uploaded by Hugo Yèche
Content may be subject to copyright.
![Illustrative example of model's behaviour for decompensation. (left) Predictions for EHRonly (blue) and cross-modal (orange) models over time compared to ground truth labels (green). (right) Cross-attention over time for the cross-modal model where EHR timesteps (x-axis) attend to clinical notes (y-axis). Both figures represent patient 12767. nursing / other nursing progress notes : pt remains [ * * name ( ni ) * * ] / vented 40 % 5ps and 5peep . awaiting arrival of one more son to add in decision process of code status and extubation . lung sounds remain clear , sx for minimal amt ' s of thick tan secretions . 02 sat ' s cont in high 90 ' s . cv : hr 80 ' s to 90 ' s nsr no ectopy . bp stable with lopressor see careview for data . temp max 101 . tylenol given . gi : foley patent draining mod amt ' s of clear yellow urine . gu : ngt clamped . npo . no stool . abd soft distended , hypoactive bowel sounds . neuro : no chg in neuro status . cont to be unresponsive , does not open eyes . moving right arm and right leg , no movement on left side . pupils 2mm react briskly . iv fluid cont at 40 / hr . pt waiting till morning to most likely be made dnr / di .](profile/Hugo-Yeche/publication/366062908/figure/fig2/AS:11431281105385317@1670383349379/Illustrative-example-of-models-behaviour-for-decompensation-left-Predictions-for.png)
Illustrative example of model's behaviour for decompensation. (left) Predictions for EHRonly (blue) and cross-modal (orange) models over time compared to ground truth labels (green). (right) Cross-attention over time for the cross-modal model where EHR timesteps (x-axis) attend to clinical notes (y-axis). Both figures represent patient 12767. nursing / other nursing progress notes : pt remains [ * * name ( ni ) * * ] / vented 40 % 5ps and 5peep . awaiting arrival of one more son to add in decision process of code status and extubation . lung sounds remain clear , sx for minimal amt ' s of thick tan secretions . 02 sat ' s cont in high 90 ' s . cv : hr 80 ' s to 90 ' s nsr no ectopy . bp stable with lopressor see careview for data . temp max 101 . tylenol given . gi : foley patent draining mod amt ' s of clear yellow urine . gu : ngt clamped . npo . no stool . abd soft distended , hypoactive bowel sounds . neuro : no chg in neuro status . cont to be unresponsive , does not open eyes . moving right arm and right leg , no movement on left side . pupils 2mm react briskly . iv fluid cont at 40 / hr . pt waiting till morning to most likely be made dnr / di .
Source publication
Understanding deep learning model behavior is critical to accepting machine learning-based decision support systems in the medical community. Previous research has shown that jointly using clinical notes with electronic health record (EHR) data improved predictive performance for patient monitoring in the intensive care unit (ICU). In this work, we...
Contexts in source publication
Context 1
... further illustrate our conclusions, we also present a second example corresponding to the patient with ID 12767. Figure 9 shows the predictions and the corresponding cross-attention from ICU 10 12 14 16 18 20 22 24 26 28 30 32 34 36 38 40 42 44 46 48 50 52 54 56 58 60 62 64 66 Timestep ( time series to text series over time. We observe that at the 63rd timestep when the multi-modal transformer's prediction deviates from the EHR-only transformer's prediction it attends to the newly available 8th note. ...
Context 2
... in the previous example, we compute the attention rollout from this particular note in Figure 9. This short note states that the patient is most likely to be made DNR/DNI. ...
Context 3
... further illustrate our conclusions, we also present a second example corresponding to the patient with ID 12767. Figure 9 shows the predictions and the corresponding cross-attention from ICU 10 12 14 16 18 20 22 24 26 28 30 32 34 36 38 40 42 44 46 48 50 52 54 56 58 60 62 64 66 Timestep ( time series to text series over time. We observe that at the 63rd timestep when the multi-modal transformer's prediction deviates from the EHR-only transformer's prediction it attends to the newly available 8th note. ...
Similar publications
Background
The demanding nature of neurosurgical practice requires a reliable system for documentation and record-keeping. The cost of electronic health record systems can limit their availability in low- and middle-income countries. That is why less expensive and easily accessible technological alternatives should be sought. In this article, we de...
