INS NYC 2024 Program

Poster	Poster Session 04 Program Schedule 02/15/2024 12:00 pm - 01:15 pm Room: Shubert Complex (Posters 1-60) Poster Session 04: Neuroimaging \| Neurostimulation/Neuromodulation \| Teleneuropsychology/Technology Final Abstract #12 Deep learning representations of the clock drawing test can improve post-operative outcome predictions Sabyasachi Bandyopadhyay, University of Florida, Gainesville, United States Ronald Ison, University of Florida, Gainesville, United States David Libon, Rowan University, Stratford, United States Patrick Tighe, University of Florida, Gainesville, United States Catherine Price, University of Florida, Gainesville, United States Parisa Rashidi, University of Florida, Gainesville, United States Category: Assessment/Psychometrics/Methods (Adult) Keyword 1: assessment Keyword 2: cognitive screening Keyword 3: dementia - Alzheimer's disease Objective: The association between pre-operative cognitive status and post-operative outcomes is critical, yet scarcely explored field. Collecting intra-operative data represents a low-cost way to evaluate long-term impacts of surgical interventions. In this project, we evaluated how pre-operative cognitive status as measured by the clock drawing test (CDT) contributed to predicting length of hospital stay, hospital charges, average pain experienced during follow-up, and one-year mortality over and above intra-operative variables, demographics, pre-operative physical status and co-morbidities. Participants and Methods: The data was collected through as part of an institutional review board (IRB) approved study at the University of Florida (UF) and UF Health Shands hospital. The original dataset consisted of 22,473 patients amongst which 6,221 patients had complete data from all three data modalities: a) pre-operative clock drawings, b) demographics and c) intra-operative variables. The clock drawing images were represented by ten constructional features discovered by a semi-supervised deep learning (DL) algorithm, previously validated to differentiate between dementia and non-dementia patients. Different machine learning (ML) models were trained using 5-fold cross validation to classify post-operative outcomes in hold-out test sets. Shapeley Additive Explanations (SHAP) analysis was used to find the most predictive features for classifying different outcomes in different surgical contexts. Results: For all outcomes, a combination of pre-operative cognition, intra-operative features and clinical characteristics yielded the best classifiers. For length of stay the best classifier gave 0.93 AUC, 0.83 F1-score, 0.82 Precision, 0.85 Sensitivity and 0.88 Specificity. For predicting hospital charges the best classifier gave 0.98 AUC, 0.93 F1-score, 0.94 Precision, 0.92 Sensitivity and 0.94 Specificity. For average pain the best classifier gave 0.82 AUC, 0.80 F1-score, 0.80 Precision, 0.80 Sensitivity and 0.73 Specificity. For mortality prediction the best classifier gave 0.71 AUC, 0.16 F1-score, 0.09 Precision, 0.61 Sensitivity and 0.71 Specificity. SHAP analysis found that duration of surgery was the most important predictor of adverse post-operative outcomes. Conclusions: This study showed that although surgical variables were the most predictive when classifying post-operative adverse outcomes, addition of preoperative clock drawing features derived from deep learning representations which are informative of the cognitive health of participants can yield stronger classifiers across multiple post-operative outcomes.

Poster

02/15/2024
12:00 pm - 01:15 pm
Room: Shubert Complex (Posters 1-60)

Poster Session 04: Neuroimaging | Neurostimulation/Neuromodulation | Teleneuropsychology/Technology

Final Abstract #12

Deep learning representations of the clock drawing test can improve post-operative outcome predictions

Sabyasachi Bandyopadhyay, University of Florida, Gainesville, United States
Ronald Ison, University of Florida, Gainesville, United States
David Libon, Rowan University, Stratford, United States
Patrick Tighe, University of Florida, Gainesville, United States
Catherine Price, University of Florida, Gainesville, United States
Parisa Rashidi, University of Florida, Gainesville, United States

Category: Assessment/Psychometrics/Methods (Adult)

Keyword 1: assessment
Keyword 2: cognitive screening
Keyword 3: dementia - Alzheimer's disease

Objective:

The association between pre-operative cognitive status and post-operative outcomes is critical, yet scarcely explored field. Collecting intra-operative data represents a low-cost way to evaluate long-term impacts of surgical interventions. In this project, we evaluated how pre-operative cognitive status as measured by the clock drawing test (CDT) contributed to predicting length of hospital stay, hospital charges, average pain experienced during follow-up, and one-year mortality over and above intra-operative variables, demographics, pre-operative physical status and co-morbidities.

Participants and Methods:

The data was collected through as part of an institutional review board (IRB) approved study at the University of Florida (UF) and UF Health Shands hospital. The original dataset consisted of 22,473 patients amongst which 6,221 patients had complete data from all three data modalities: a) pre-operative clock drawings, b) demographics and c) intra-operative variables. The clock drawing images were represented by ten constructional features discovered by a semi-supervised deep learning (DL) algorithm, previously validated to differentiate between dementia and non-dementia patients. Different machine learning (ML) models were trained using 5-fold cross validation to classify post-operative outcomes in hold-out test sets. Shapeley Additive Explanations (SHAP) analysis was used to find the most predictive features for classifying different outcomes in different surgical contexts.

Results:

For all outcomes, a combination of pre-operative cognition, intra-operative features and clinical characteristics yielded the best classifiers. For length of stay the best classifier gave 0.93 AUC, 0.83 F1-score, 0.82 Precision, 0.85 Sensitivity and 0.88 Specificity. For predicting hospital charges the best classifier gave 0.98 AUC, 0.93 F1-score, 0.94 Precision, 0.92 Sensitivity and 0.94 Specificity. For average pain the best classifier gave 0.82 AUC, 0.80 F1-score, 0.80 Precision, 0.80 Sensitivity and 0.73 Specificity. For mortality prediction the best classifier gave 0.71 AUC, 0.16 F1-score, 0.09 Precision, 0.61 Sensitivity and 0.71 Specificity. SHAP analysis found that duration of surgery was the most important predictor of adverse post-operative outcomes.

Conclusions:

This study showed that although surgical variables were the most predictive when classifying post-operative adverse outcomes, addition of preoperative clock drawing features derived from deep learning representations which are informative of the cognitive health of participants can yield stronger classifiers across multiple post-operative outcomes.