Safe and Explainable AI

NCT06694181 | Recruiting

• 1 other identifier: 34124
• Study Type: observational
• Target: 300,000
• Locations: 1 country
• Sites: 1

Brief Summary

While current AI technology is suitable for automating some repetitive clinical tasks, technical challenges remain in solving critical and gainful problems in the domains of patient and disease management. The proposed research seeks to address issues in medical AI, such as integrating medical knowledge effectively, making AI recommendations explainable to clinicians, and establishing safety guarantees.

Conditions

Artifical Inteligence; Cardiology; Breast Cancer; Sepsis

Outcome Measures

Primary Outcomes (3)

• Neurosymbolic Learning Algorithms

  Develop and evaluate novel algorithms for training neurosymbolic models. We will develop data- and compute-efficient algorithms for end-to-end training of neurosymbolic models. This task will reduce the burden on clinician experts to provide fine-grained labels on voluminous EHR data.

  Prototype and develop new learning algorithms; 18 months. Benchmark and evaluate the learning algorithms; 24 months. Publish research results; 24 months

• Explanation Methods

  We will develop new explainable AI techniques that come with verifiable guarantees. These guarantees will enable trust and transparency in AI at a fundamental level.

  Prototype and develop new explanation algorithms; 18 months. Derive certified guarantees for explanations; 18 months. Benchmark and evaluate the explanation algorithms; 24 months. Extend certificates to new properties and tasks; 30 months. Publ

• Methods for Safety Guarantees

  We will develop new algorithms that can scalably extract complex logical rules governing safety within the data that have statistical guarantees. These techniques will be rooted in statistical analysis and assist users in identifying out of distribution data and detecting anomalies.

  Prototype and develop new rule learning algorithms; 30 months. Scale rule learning algorithms to larger data settings; 36 months. Incorporate new primitives to express complex rules; 36 months. Implement rule learning algorithms on baseline tasks

Study Arms (3)

Cardiology

The primary objective in this clinical case scenario is to evaluate an ML model utilizing real-time cardiac telemetry, as well as other clinical, demographic, and imaging structured data sources, among hospitalized, intensive care unit (ICU) patients to predict impending inhospital cardiac arrest, identify potentially reversible causes of cardiac arrest, and predict which patients may have impending cardiac arrest due to shockable rhythms i.e. ventricular tachycardia (VT) or ventricular fibrillation (VF).

Other: AI-PERSONALIZED CLINICAL DECISION SUPPORT

Oncology - Breast Cancer

The primary objective in this clinical case scenario is to evaluate an ML model utilizing structured and unstructured data from clinical, demographic, and tumor molecular and germline sequencing, among outpatients with cancer, to predict short-term mortality and/or symptom decline. The model for prediction to treatment response in breast cancer patients will be compared with two prognostic tools: 1) Conversation Connect, a previously validated machine learning mortality prediction tool that has been used at the University of Pennsylvania for routine clinical decision support, and 2) the Elixhauser Comorbidity Index, a comorbidity-based prognostic index used commonly in research and risk-adjustment.

Other: AI-PERSONALIZED CLINICAL DECISION SUPPORT

Sepsis

The primary objective in this clinical case scenario is to develop and evaluate an ML model that utilizes multidmodal clinical data (e.g., structured EHR data such as demographics, laboratory test results, and vital signs; unstructured EHR data including the text of clinical encounter notes and, where available, waveforms from real-time cardiac, hemodynamic, and respiratory monitoring devices) to predict the need for initiation of broad-spectrum antimicrobial therapy for hospitalized patients with sepsis. With a focus on implementable and explainable AI, we will produce well calibrated predictions that are also clinically meaningful at the bedside to aid real-time decision-making about diagnosis and treatment initiation. The model for timely diagnosis and intervention in sepsis will be compared with widely used commercial and open-source sepsis prediction models.

Other: AI-PERSONALIZED CLINICAL DECISION SUPPORT

Interventions

AI-PERSONALIZED CLINICAL DECISION SUPPORT OTHER

AI-ENABLED DECISION MAKING FOR PERSONALIZED CLINICAL DECISION SUPPORT

Cardiology; Oncology - Breast Cancer; Sepsis

Eligibility Criteria

• Age: 18 Years+
• Sex: all
• Healthy Volunteers: No
• Age Groups: Adult (18-64), Older Adult (65+)
• Sampling Method: Non-Probability Sample

Study Population

The study will look at EMR data from cardiac, oncology, and patients at risk for acquiring sepsis. Each group has clearly defined inclusion-exclusion criteria. For the machine learning model to predict ventricular arrhythmias and in-hospital cardiac arrest, we will construct patient cohorts consisting of adult (i.e., 18 years of age or older) patients admitted to Sickbay-accessible bed within a Penn Medicine hospital from 2020 to the present as well as patients identified as having experienced cardiac arrest using EHR records.

You may qualify if:

• Cardiology 18 years of age and older, admitted to any of the Penn Medicine hospitals from 2017 to the present. Sepsis 18 years of age at the time of presentation to an emergency department or admission to any Penn Medicine hospital from July 1, 2017, onward will be eligible as this represents the population at risk for acquiring sepsis Oncology 18 years of age and older with a diagnosis of invasive breast cancer (Stage 1-4) in the Penn Cancer registry

Sponsors & Collaborators

• Abramson Cancer Center at Penn Medicine: lead

Study Sites (1)

Hospital of the University of Pennsylvania

Philadelphia, Pennsylvania, 19104, United States

RECRUITING

MeSH Terms

Conditions

Breast Neoplasms; Sepsis

Condition Hierarchy (Ancestors)

Neoplasms by Site; Neoplasms; Breast Diseases; Skin Diseases; Skin and Connective Tissue Diseases; Infections; Systemic Inflammatory Response Syndrome; Inflammation; Pathologic Processes; Pathological Conditions, Signs and Symptoms

Central Study Contacts

Haideliza Soto Calderon

CONTACT

215-220-9425; haideliza.soto-calderon@pennmedicine.upenn.edu

Nicholas Bishop

CONTACT

nicholas.bishop@pennmedicine.upenn.edu

Study Design

• Study Type: observational
• Observational Model: CASE ONLY
• Time Perspective: RETROSPECTIVE
• Sponsor Type: OTHER
• Responsible Party: SPONSOR

Study Record Dates

• First Submitted: November 15, 2024
• First Posted: November 19, 2024
• Study Start: November 29, 2025
• Primary Completion (Estimated): November 1, 2028
• Study Completion (Estimated): November 1, 2028
• Last Updated: February 25, 2026

Record last verified: 2026-02

Data Sharing

• IPD Sharing: Will not share

Locations

US (1)