Tempus AI, Inc. · NASDAQ:TEM
Tempus AI — Technical Presentation (November 2022)
19 slides from Tempus AI's technical presentation, November 2022
Slide 1: EHR-based Machine Learning Model Predicts Drug-induced QT Prolongation
EHR-based Machine Learning Model Predicts Drug-induced QT Prolongation With Superior Performance Compared To Clinical Risk Predictors Linyuan Jing, PhD,1,2 Thomas Morland, MD,1 Christopher Kelsey,1 Sushravya Raghunath, PhD,2 John Pfeifer, MD,2 Jeffrey Ruhl,1 Brandon Fornwalt, MD, PhD,2 Christopher H
Slide 2: Disclosures
Disclosures • Funding • Geisinger receives funding from Tempus for ongoing development of predictive modeling technology and commercialization • None of the Geisinger authors have ownership interests in any of the intellectual property resulting from the partnership. However, Dr. Fornwalt, Dr. Raghu
Slide 3: What is Drug-induced Long QTc?
What is Drug-induced Long QTc? • Drug-induced long QTc is an iatrogenic delay in ventricular repolarization which can lead to the life-threatening arrhythmia, torsades de pointes (TDP) • >200 drugs are QTc prolongers • Antiarrhythmics • Antibiotics • Antipsychotics • Antidepressants • Cancer therapi
Slide 4: Current Solutions
Current Solutions Tisdale Score Optimized RISQ-PATH Score • Predicts QTc >500 ms or >60 ms • Broader application compared to baseline • Applicable to outpatient • Prospective, observational study • Moderate performance: AUC=0.772 of 1,200 patients in cardiac critical care unit • Good performance: AU
Slide 5: Potential for a Machine Learning Approach
Potential for a Machine Learning Approach Machine learning has demonstrated high accuracy (e.g., AUROC >0.900) on targets related to QTc but has not been used to produce a general clinical model with high accuracy
Slide 6: Design Criteria
Design Criteria • Highly accurate across a variety of clinical settings and scenarios • Derived and validated utilizing data commonly available in the electronic health record (EHR) at the time of medical decision- making • Assessment of drug-induced long QTc risk in general population as opposed to
Slide 7: Study Design
Study Design Structure EHR features Demographics Lab tests High risk of long QTc Clean 1 Vital Signs Medications XGBoost Scale Classifier Impute ECG measures ECG patterns Risk Score 12-lead, 10-second voltage data x4 64, RELU Dense Age, Sex (n, 32, 10) Concatenation Concatenation (2461, 32, 80) (n,
Slide 8: Study Population
Study Population Index -3 year -1 year Time 0 1 year Pre-drug start QTc drug start QTc drug end Epic EHR features Target ECG XGB set Ø Endpoint: QTc > 500 ms Baseline ECG features Prediction window from target on-drug ECG ECG Baseline ECG traces DNN set XGB dataset ECG DNN dataset N patients 345,371
Slide 9: Both XGBoost and DNN Trace Models Show Comparable Performance
Both XGBoost and DNN Trace Models Show Comparable Performance 86.4% 86.9% 87.4% • Dataset: • N = 182,448, Structured EHR features • 7.7% long QTc (>500 ms) ECG trace + age/sex • No significant improvement in composite model EHR features + DNN risk score All models were evaluated by 5-fold cross-vali
Slide 10: Machine Learning Shows Superior Predicting Power to Tisdale and RISQ-PATH
Machine Learning Shows Superior Predicting Power to Tisdale and RISQ-PATH AUC: 0.86 AUC: 0.86 AUC: 0.77 AUC: 0.70 N=110,588, 8.8% long QTc N=345,371, 5.7% long QTc Tisdale scores were only available in patients with ECGs within 24h of drug start
Slide 11: Machine Learning Shows Superior Performance in All Clinical Settings
Machine Learning Shows Superior Performance in All Clinical Settings Setting Data size % Long QTc Total 110,588 8.8 Office Visit 32,271 3.3 Admission 62,528 13 ED 15,759 3.6 • Tisdale scores were predominantly calculated in hospital admissions • In contrast, 75% of the overall dataset were from outp
Slide 12: Machine Learning Model Has Better PPV
Machine Learning Model Has Better PPV • When matching number of patients predicted as high-risk for long QTc, XGBoost model shows superior PPV (>50%) as compared to Tisdale and RISQ-PATH % Long % Predicted Sensitivity Specificity PPV NPV QTc High-risk 91.8 3.1 9.8 97.6 28.3 Tisdale [90.9, 8.8 [1.7,
Slide 13: Conclusions
Conclusions • Machine learning models, either using structured EHR features or using ECG traces alone, can predict drug-induced QTc prolongation at medication initiation in a general clinical population with high accuracy Ø Composite model using both EHR features and DNN risk score shows marginal im
Slide 14: Limitations
Limitations • Model trained on all QTc drugs: variable QTc risks and different mechanisms of action • Model was evaluated on commonly used, individual drugs but performance was unreliable due to small sample size for single medication • Model endpoint QTc >500 ms • Rational: Ø QTc > 500 ms can be ha
Slide 15: Acknowledgements
Acknowledgements • Geisinger and Tempus teams • Tempus Laboratories funding • Thank you! [email protected]
Slide 16: Appendix
Appendix Geisinger ‘TEMPUS
Slide 17: Machine Learning Model Shows Better Metrics than Original Tisdale/RISQ-PATH Studies
Machine Learning Model Shows Better Metrics than Original Tisdale/RISQ-PATH Studies • When matching sensitivities originally reported by Tisdale and RISQ-PATH, XGBoost model shows better specificity. % Long % Predicted Sensitivity Specificity QTc High-risk Original Tisdale 30.7 74 77 Cohort 5.7 21.6
Slide 18: Patient Characteristics
Patient Characteristics Variable Median[IQR] or % Variable Median[IQR] or % N patients 345,371 Heart failure 13% Long QTc 5.7% Diabetes 21% Age, yrs 62.2 [48.4, 74.4] Hypertension 49% Male sex 45.1% Myocardial Infarction 11% White race 95.7% QTc, ms 440 [421, 462] BMI, kg/cm2 29.5 [25.1, 35] Max QTc
Slide 19: Machine Learning Shows Superior Performance in All Clinical Settings
Machine Learning Shows Superior Performance in All Clinical Settings Setting Data size % Long QTc Tisdale subset Office Visit 32,271 3.3 Admission 62,528 13 ED 15,759 3.6 Overall dataset Office Visit 262,294 4.1 Admission 66,137 12.9 ED 16,940 3.7