Aligning SDOH Data with Patient Safety and AI Innovation

Social determinants of health (SDOH), non-medical factors such as where people are born, grow, live, go to school, work, and age, need to be an integral part of the patient safety conversation. The problem: While SDOH have become a common part of healthcare’s vernacular, there has been relatively little attention paid to how these factors affect patient safety, according to an article published in the Journal of Patient Safety.

This is unfortunate, as SDOH have the potential to both positively and negatively influence health and need to be considered when addressing patient safety, according to studies cited in a Journal of Patient Safety and Risk Management article. To do this effectively, healthcare organizations must invest in the collection and integration of SDOH data that is structured, comprehensive, and reflective of patient realities.

Why SDOH Data Matters

SDOH data refers to structured or semi-structured information collected about an individual’s social and economic conditions that influence their health. It may be:

• Self-reported via surveys (e.g., PRAPARE)
• Extracted from clinical notes using natural language processing
• Linked from external datasets (e.g., census data, housing statistics)

Common data points include housing stability, food security, access to transportation, education level, employment status, social isolation, and health literacy.

Structured capture of this information is crucial. Without it, patient safety goals and AI systems are limited by blind spots that fail to account for the real-world barriers patients face. The future of patient-centered care depends on fully recognizing and acting on SDOH data.

How SDOH Data Impacts Patient Safety Goals

According to several studies cited in the article Social Determinants of Patient Safety: A Bridge to Better Quality of Care, SDOH can significantly impact patient safety in various ways. These factors can create barriers that increase the likelihood of medical errors or adverse events. For example:

• Low health literacy can lead to unsafe medication use. A patient with diabetes who does not fully understand how to manage their insulin regimen may struggle to maintain blood sugar levels within the narrow therapeutic range, increasing the risk of complications.
• Language barriers may prevent patients from fully understanding medical instructions, which can lead to misunderstandings about medications, procedures, or care plans, raising the risk of preventable safety events.
• Lack of reliable transportation can limit access to critical follow-up care. For instance, a patient prescribed blood thinners for atrial fibrillation who misses regular monitoring appointments due to transportation issues may face heightened risks of stroke or excessive bleeding.

Each of these scenarios directly impacts organizational patient safety goals, including safe medication use, clear communication, and follow-up care. Capturing these issues as SDOH data is essential for designing risk-aware safety strategies and targeted care interventions.

Considering Lower Health Literacy and Discrimination as Factors in SDOH Data

A look at how risk managers and other healthcare professionals can improve equity and patient safety by addressing health literacy and discrimination illustrates exactly why SDOH are so important. A large body of evidence suggests that healthcare organizations must address literacy and discrimination through a patient safety lens, according to a review of studies published in the Journal of Healthcare Risk Management.

Not surprisingly, because SDOH is so important, these elements also must be included when developing AI and other technology solutions designed to help improve patient safety, according to the review. AI can provide decision support by identifying patients at high risk of harm and by guiding prevention and early intervention strategies.

Lower health literacy, for instance, is a critical SDOH data point that is often underreported yet directly tied to adverse outcomes. By tracking this data, healthcare providers can better anticipate misunderstandings in patient instruction and adjust care delivery accordingly. Likewise, recording instances or risks of discrimination enables organizations to address structural inequities that may otherwise remain invisible in clinical decision support tools and machine learning applications.

Exploring SDOH Data Shortcomings in AI

The problem, according to a presentation from the National Institute of Nursing Research, is that:

• Many AI systems lack representative, diverse data and fail to examine how the inclusion of SDOH may affect results and model interpretation.
• Some AI technologies do not include SDOH even though research shows that including SDOH significantly improves model fit, outcomes, and predictions.
• When SDOH are utilized in AI algorithms, data inputs rarely go beyond including individual-level SDOH proxies such as race/ethnicity, gender, or age, which do not capture the wider set of structural factors shaping the conditions of daily life.

The limitations of such models reinforce the urgency of capturing meaningful SDOH data at both the individual and population levels. Without this, AI systems may perpetuate bias, exacerbate disparities, or make incomplete and potentially harmful predictions. Expanding the scope of SDOH data used in model training and validation is necessary for building truly responsible AI in healthcare.

Integrating SDOH Data into AI

In a question-and-answer article published in Healthcare IT News, Jim Watson, MD, president and CEO of Genesis Physicians Group, described how to integrate SDOH into AI systems.

According to Watson, while specific populations’ social needs can be identified using publicly available social data related to a person’s address (census tract and/or ZIP code level data), this basic data is simply not enough. Instead, when creating AI and machine learning models, it’s important to:

• Collect individual patient-reported data around expressed social needs that often create barriers to accessing healthcare, as well as healthcare outcomes.
• Include a patient’s clinical data (for example, utilization, costs, and pharmaceutical utilization) and mix it with both population-based social determinant data and individually reported social need data to create a more complete risk profile stratification process for a specific population.
• Use the data to stratify risk. With machine learning technology, data scientists can risk-stratify the population, placing patients with higher burdens of social risk impacting their health access and outcomes at the top, and those with less burden toward the bottom.
• Reuse the data. As organizations and provider networks intervene on identified social needs, changes in clinical outcomes and social needs can be used in a feedback loop to retrain the machine learning algorithm. This helps the model become more precise in determining which social need intervention may have the highest likelihood of producing the greatest positive impact, improving efficiency for the organization.
• Train custom AI/machine learning models using data from the specific population they will serve, rather than relying on off-the-shelf models built on general datasets. This approach is critical because predictive accuracy declines when models are applied to populations that differ from the ones they were trained on.

Custom AI solutions that integrate SDOH data are essential for organizations committed to precision, fairness, and long-term effectiveness in care delivery. Here is a comparison of off-the-shelf versus custom AI models in the context of implementing AI in healthcare:

Feature	Off-the-Shelf AI Model	Custom AI Model Trained with SDOH Data
Data Source	Generic datasets, often limited SDOH inclusion	Specific patient population, including SDOH
Risk Stratification Accuracy	Moderate, may miss key social risk indicators	High, driven by relevant social and clinical data
Bias Risk	Higher risk due to lack of contextual factors	Lower risk when trained on diverse, local data
Use Case Fit	Broad applications, limited personalization	Tailored predictions for a specific population
Feedback & Retraining Capability	Often static and fixed	Designed for ongoing feedback and optimization

Addressing SDOH Data Deficiencies

Off-the-shelf models often are trained on certain data types or data sources. As such, predictive accuracy may drop. To alleviate this problem, healthcare organizations should leverage other data. Genesis Physicians Group, for example, conducts individual interviews or surveys around SDOH and social needs that are highly connected to the risk of future adverse events that aren’t easily incorporated into off-the-shelf predictive models.

These alternative data sources can address a variety of social factors, from food insecurity to social isolation, and support responsible AI in healthcare by improving accuracy without introducing bias.

By incorporating diverse, patient-specific SDOH data, AI models become more precise and more ethical. Ultimately, responsible integration of this data supports healthcare systems in not only anticipating risks but also tailoring interventions that advance equity and improve patient outcomes.