User login
Researchers found SDOH are risk factors for suicide among U.S. veterans and NLP can be leveraged to extract SDOH information from unstructured data in the EHR.
“Since SDOH is overwhelmingly described in EHR notes, the importance of NLP-extracted SDOH can be very significant, meaning that NLP can be used as an effective method for epidemiological and public health study,” senior investigator Hong Yu, PhD, from Miner School of Information and Computer Sciences, University of Massachusetts Lowell, told this news organization.
Although the study was conducted among U.S. veterans, the results likely hold for the general population as well.
“The NLP methods are generalizable. The SDOH categories are generalizable. There may be some variations in terms of the strength of associations in NLP-extracted SDOH and suicide death, but the overall findings are generalizable,” Dr. Yu said.
The study was published online JAMA Network Open.
Improved risk assessment
SDOH, which include factors such as socioeconomic status, access to healthy food, education, housing, and physical environment, are strong predictors of suicidal behaviors.
Several studies have identified a range of common risk factors for suicide using International Classification of Diseases (ICD) codes and other “structured” data from the EHR. However, the use of unstructured EHR data from clinician notes has received little attention in investigating potential associations between suicide and SDOH.
Using the large Veterans Health Administration EHR system, the researchers determined associations between veterans’ death by suicide and recent SDOH, identified using both structured data (ICD-10 codes and Veterans Health Administration stop codes) and unstructured data (NLP-processed clinical notes).
Participants included 8,821 veterans who committed suicide and 35,284 matched controls. The cohort was mostly male (96%) and White (79%). The mean age was 58 years.
The NLP-extracted SDOH were social isolation, job or financial insecurity, housing instability, legal problems, violence, barriers to care, transition of care, and food insecurity.
All of these unstructured clinical notes on SDOH were significantly associated with increased risk for death by suicide.
Legal problems had the largest estimated effect size, more than twice the risk of those with no exposure (adjusted odds ratio 2.62; 95% confidence interval, 2.38-2.89), followed by violence (aOR, 2.34; 95% CI, 2.17-2.52) and social isolation (aOR, 1.94; 95% CI, 1.83-2.06).
Similarly, all of the structured SDOH – social or family problems, employment or financial problems, housing instability, legal problems, violence, and nonspecific psychosocial needs – also showed significant associations with increased risk for suicide death, once again, with legal problems linked to the highest risk (aOR, 2.63; 95% CI, 2.37-2.91).
When combining the structured and NLP-extracted unstructured data, the top three risk factors for death by suicide were legal problems (aOR, 2.66; 95% CI 2.46-2.89), violence (aOR, 2.12; 95% CI, 1.98-2.27), and nonspecific psychosocial needs (aOR, 2.07; 95% CI, 1.92-2.23).
“To our knowledge, this the first large-scale study to implement and use an NLP system to extract SDOH information from unstructured EHR data,” the researchers write.
“We strongly believe that analyzing all available SDOH information, including those contained in clinical notes, can help develop a better system for risk assessment and suicide prevention. However, more studies are required to investigate ways of seamlessly incorporating SDOHs into existing health care systems,” they conclude.
Dr. Yu said it’s also important to note that their NLP system is built upon “the most advanced deep-learning technologies and therefore is more generalizable than most existing work that mainly used rule-based approaches or traditional machine learning for identifying social determinants of health.”
In an accompanying editorial, Ishanu Chattopadhyay, PhD, of the University of Chicago, said this suggests that unstructured clinical notes “may efficiently identify at-risk individuals even when structured data on the relevant variables are missing or incomplete.”
This work may provide “the foundation for addressing the key hurdles in enacting efficient universal assessment for suicide risk among the veterans and perhaps in the general population,” Dr. Chattopadhyay added.
This research was funded by a grant from the National Institute of Mental Health. The study authors and editorialist report no relevant financial relationships.
A version of this article originally appeared on Medscape.com.
Researchers found SDOH are risk factors for suicide among U.S. veterans and NLP can be leveraged to extract SDOH information from unstructured data in the EHR.
“Since SDOH is overwhelmingly described in EHR notes, the importance of NLP-extracted SDOH can be very significant, meaning that NLP can be used as an effective method for epidemiological and public health study,” senior investigator Hong Yu, PhD, from Miner School of Information and Computer Sciences, University of Massachusetts Lowell, told this news organization.
Although the study was conducted among U.S. veterans, the results likely hold for the general population as well.
“The NLP methods are generalizable. The SDOH categories are generalizable. There may be some variations in terms of the strength of associations in NLP-extracted SDOH and suicide death, but the overall findings are generalizable,” Dr. Yu said.
The study was published online JAMA Network Open.
Improved risk assessment
SDOH, which include factors such as socioeconomic status, access to healthy food, education, housing, and physical environment, are strong predictors of suicidal behaviors.
Several studies have identified a range of common risk factors for suicide using International Classification of Diseases (ICD) codes and other “structured” data from the EHR. However, the use of unstructured EHR data from clinician notes has received little attention in investigating potential associations between suicide and SDOH.
Using the large Veterans Health Administration EHR system, the researchers determined associations between veterans’ death by suicide and recent SDOH, identified using both structured data (ICD-10 codes and Veterans Health Administration stop codes) and unstructured data (NLP-processed clinical notes).
Participants included 8,821 veterans who committed suicide and 35,284 matched controls. The cohort was mostly male (96%) and White (79%). The mean age was 58 years.
The NLP-extracted SDOH were social isolation, job or financial insecurity, housing instability, legal problems, violence, barriers to care, transition of care, and food insecurity.
All of these unstructured clinical notes on SDOH were significantly associated with increased risk for death by suicide.
Legal problems had the largest estimated effect size, more than twice the risk of those with no exposure (adjusted odds ratio 2.62; 95% confidence interval, 2.38-2.89), followed by violence (aOR, 2.34; 95% CI, 2.17-2.52) and social isolation (aOR, 1.94; 95% CI, 1.83-2.06).
Similarly, all of the structured SDOH – social or family problems, employment or financial problems, housing instability, legal problems, violence, and nonspecific psychosocial needs – also showed significant associations with increased risk for suicide death, once again, with legal problems linked to the highest risk (aOR, 2.63; 95% CI, 2.37-2.91).
When combining the structured and NLP-extracted unstructured data, the top three risk factors for death by suicide were legal problems (aOR, 2.66; 95% CI 2.46-2.89), violence (aOR, 2.12; 95% CI, 1.98-2.27), and nonspecific psychosocial needs (aOR, 2.07; 95% CI, 1.92-2.23).
“To our knowledge, this the first large-scale study to implement and use an NLP system to extract SDOH information from unstructured EHR data,” the researchers write.
“We strongly believe that analyzing all available SDOH information, including those contained in clinical notes, can help develop a better system for risk assessment and suicide prevention. However, more studies are required to investigate ways of seamlessly incorporating SDOHs into existing health care systems,” they conclude.
Dr. Yu said it’s also important to note that their NLP system is built upon “the most advanced deep-learning technologies and therefore is more generalizable than most existing work that mainly used rule-based approaches or traditional machine learning for identifying social determinants of health.”
In an accompanying editorial, Ishanu Chattopadhyay, PhD, of the University of Chicago, said this suggests that unstructured clinical notes “may efficiently identify at-risk individuals even when structured data on the relevant variables are missing or incomplete.”
This work may provide “the foundation for addressing the key hurdles in enacting efficient universal assessment for suicide risk among the veterans and perhaps in the general population,” Dr. Chattopadhyay added.
This research was funded by a grant from the National Institute of Mental Health. The study authors and editorialist report no relevant financial relationships.
A version of this article originally appeared on Medscape.com.
Researchers found SDOH are risk factors for suicide among U.S. veterans and NLP can be leveraged to extract SDOH information from unstructured data in the EHR.
“Since SDOH is overwhelmingly described in EHR notes, the importance of NLP-extracted SDOH can be very significant, meaning that NLP can be used as an effective method for epidemiological and public health study,” senior investigator Hong Yu, PhD, from Miner School of Information and Computer Sciences, University of Massachusetts Lowell, told this news organization.
Although the study was conducted among U.S. veterans, the results likely hold for the general population as well.
“The NLP methods are generalizable. The SDOH categories are generalizable. There may be some variations in terms of the strength of associations in NLP-extracted SDOH and suicide death, but the overall findings are generalizable,” Dr. Yu said.
The study was published online JAMA Network Open.
Improved risk assessment
SDOH, which include factors such as socioeconomic status, access to healthy food, education, housing, and physical environment, are strong predictors of suicidal behaviors.
Several studies have identified a range of common risk factors for suicide using International Classification of Diseases (ICD) codes and other “structured” data from the EHR. However, the use of unstructured EHR data from clinician notes has received little attention in investigating potential associations between suicide and SDOH.
Using the large Veterans Health Administration EHR system, the researchers determined associations between veterans’ death by suicide and recent SDOH, identified using both structured data (ICD-10 codes and Veterans Health Administration stop codes) and unstructured data (NLP-processed clinical notes).
Participants included 8,821 veterans who committed suicide and 35,284 matched controls. The cohort was mostly male (96%) and White (79%). The mean age was 58 years.
The NLP-extracted SDOH were social isolation, job or financial insecurity, housing instability, legal problems, violence, barriers to care, transition of care, and food insecurity.
All of these unstructured clinical notes on SDOH were significantly associated with increased risk for death by suicide.
Legal problems had the largest estimated effect size, more than twice the risk of those with no exposure (adjusted odds ratio 2.62; 95% confidence interval, 2.38-2.89), followed by violence (aOR, 2.34; 95% CI, 2.17-2.52) and social isolation (aOR, 1.94; 95% CI, 1.83-2.06).
Similarly, all of the structured SDOH – social or family problems, employment or financial problems, housing instability, legal problems, violence, and nonspecific psychosocial needs – also showed significant associations with increased risk for suicide death, once again, with legal problems linked to the highest risk (aOR, 2.63; 95% CI, 2.37-2.91).
When combining the structured and NLP-extracted unstructured data, the top three risk factors for death by suicide were legal problems (aOR, 2.66; 95% CI 2.46-2.89), violence (aOR, 2.12; 95% CI, 1.98-2.27), and nonspecific psychosocial needs (aOR, 2.07; 95% CI, 1.92-2.23).
“To our knowledge, this the first large-scale study to implement and use an NLP system to extract SDOH information from unstructured EHR data,” the researchers write.
“We strongly believe that analyzing all available SDOH information, including those contained in clinical notes, can help develop a better system for risk assessment and suicide prevention. However, more studies are required to investigate ways of seamlessly incorporating SDOHs into existing health care systems,” they conclude.
Dr. Yu said it’s also important to note that their NLP system is built upon “the most advanced deep-learning technologies and therefore is more generalizable than most existing work that mainly used rule-based approaches or traditional machine learning for identifying social determinants of health.”
In an accompanying editorial, Ishanu Chattopadhyay, PhD, of the University of Chicago, said this suggests that unstructured clinical notes “may efficiently identify at-risk individuals even when structured data on the relevant variables are missing or incomplete.”
This work may provide “the foundation for addressing the key hurdles in enacting efficient universal assessment for suicide risk among the veterans and perhaps in the general population,” Dr. Chattopadhyay added.
This research was funded by a grant from the National Institute of Mental Health. The study authors and editorialist report no relevant financial relationships.
A version of this article originally appeared on Medscape.com.
FROM JAMA NETWORK OPEN