Page 124 - AIH-1-2

Artificial Intelligence in Health SDoH in clinical narratives

1. Introduction in clinical decision-making. Within the domain of medical
literature, clinical case reports serve as a reflection of the
Social determinants of health (SDoH) are fundamental priorities and perspectives of health-care professionals. The
conditions that underpin the health disparities experienced choices they make in detailing specific patient information
by individuals globally. These determinants are the — what they choose to include or exclude—offer insights
circumstances in which people are born, grow, work, and into what they deem significant or irrelevant. As such,
live, and they encompass factors such as socioeconomic the inclusion or omission of SDoH in these published
status, housing, food security, and exposure to violence reports can act as a barometer of their importance within
1,2
or stress. Notably, these conditions have been proven to the health-care community. By analyzing the frequency
shape health outcomes to such an extent that up to 40% of and context of SDoH mentions in these clinical cases, one
health outcomes are attributed to SDoH challenges. 3,4
can gauge the weight and significance attributed to these
Significantly, SDoH not only impacts health outcomes factors by health-care professionals when communicating
but also has discernible effects on health-care utilization. notable clinical findings to a wider scientific audience.
For instance, unmet social needs, a facet of SDoH, have been
5
tied to clinical outcomes such as uncontrolled diabetes, Natural language processing (NLP) has become an
hypertension, and increased hospital readmissions for indispensable tool in the medical domain, revolutionizing
6
the extraction and analysis of complex data from clinical
heart failure. There is also evidence suggesting that 19,20
7
moving from a high-poverty neighborhood to one with texts and patient records. Recent publications highlight
lower poverty levels can lead to reductions in conditions the crucial role of NLP in identifying, categorizing, and
such as extreme obesity and diabetes, emphasizing the role analyzing health-related information from unstructured
of environmental factors on health. 8 content as clinical narratives. The advancements in
NLP technologies, such as context-aware models like
Given the undeniable influence of SDoH on health, Bidirectional Encoder Representations from Transformers
there have been initiatives to incorporate SDoH screening (BERT) and BioBERT, have dramatically enhanced
21
22
into health-care delivery, with proposals to standardize the our ability to process vast datasets, thereby transforming
methods for capturing this information in electronic health traditional health-care data analysis methods. 23-26 These
records (EHRs). Advocates believe that documenting innovations offer deeper insights into the prevalence
9
SDoH systematically at the point of care would bolster and impact of SDoH, previously obscured in clinical
the identification of patients’ risk factors and streamline documentation. For instance, research has demonstrated
27
referrals to social services, fostering a more holistic that NLP-based systems can identify clinical events with
approach to patient care. 10,11 significantly higher precision and sensitivity compared to
However, the current reality paints a different picture. traditional methods. One study demonstrated that an NLP
Despite the evident significance of SDoH, they remain system identified approximately four times as many clinical
underrepresented in clinical documentation. Recent events as standard approaches, with a positive predictive
studies have indicated that a mere 2% of patients visiting value (PPV) of 74%, a stark improvement over the 31% PPV
28
community health centers had at least one documented of methods relying solely on diagnostic codes. In another
12
SDoH, a figure that was confirmed by the analysis of the study, the precision of selected cases increased from 46%
ICD10 codes in different studies. 13,14 Moreover, another to 86% after incorporating NLP methods that followed
study examining over a million unique patient EHRs structured-based case selection with a sensitivity of 77%.
29
found that only a small percentage contained mentions These examples highlight the transformative impact of
of social isolation, housing issues, or financial strain, NLP in enhancing the detection and characterization
15
a finding that has been replicated in other studies. of SDoH and clinical events from medical narratives,
16
However, other analyses conducted in the primary care enabling a more nuanced and comprehensive analysis of
context have reported slightly higher proportions i.e., health-care data.
7% of patients with SDoH documented in Spain and Our study utilizes advanced NLP technology to meet
17
4% to 18% in the United States (US). These findings the need for improved documentation and understanding
18
indicate that utilizing EHRs for SDoH documentation is of SDoH in clinical settings. We investigated factors
insufficient, and a systemic approach involving education, influencing the mention of SDoH in publicly available
policy redesign, and incentives might be necessary to boost clinical case reports and how this knowledge could inform
documentation. 9 the development of more effective policies for SDoH
These findings are concerning as a discrepancy in SDoH reporting. In addition, our analysis identified potential
documentation could be indicative of a broader oversight stereotypes or discrimination in artificial intelligence (AI)

Volume 1 Issue 2 (2024) 118 doi: 10.36922/aih.2737