Several recent studies have identified that a recent diagnosis of dementia may be associated with increased risk of suicide ideation and behavior (Günak etal., Reference Günak, Barnes, Yaffe, Li and Byers2021; Schmutte etal., Reference Schmutte, Olfson, Maust, Xie and Marcus2021). It is unclear why a new diagnosis of dementia confers such risk, with possible explanations including distress associated with receiving a diagnosis with the implications that worsen with time and other factors such as psychiatric comorbidities that accompany the early phases of the illness. It may also be that preserved cognition relative to latter stages of dementia could, in some form, translate to increased likelihood of acting on suicidal thoughts by way of increased ability to plan and carry out suicidal behaviors. Thus, understanding the phenomenology of suicide in newly diagnosed dementia is a pressing area of research. Also, since the majority of data on diagnoses, dementia, and suicide risk derive from cohort studies, it is difficult to understand how suicide and dementia converge in the clinic setting. In this issue of International Psychogeriatrics, Naismith etal., (Reference Naismith, Howard, Stewart, Pitman and Mueller2022) investigate whether natural language processing (NLP) applied to free text in the Electronic Health Record (EHR) could add to the understanding of dementia and suicide risk.

Naismith etal.’s (Reference Naismith, Howard, Stewart, Pitman and Mueller2022) study is notable both for its results as well as the means by which the results were derived. Key takeaways include that suicide ideation appears to be more commonly included in records of people with Lewy Body subtype and those with depression than in other subtypes such as vascular dementia. Interestingly, putative risk factors for ideation within dementia, such as hallucinations and agitated behavior, were not related to suicide ideation. From a methodological standpoint, the study provides an example of how NLP can be applied to timely research questions using the coded and free- text fields of the EHR. The free-text field in EHR is a “catch all” for information that cannot otherwise be captured by structured and coded fields. In the EHR, the free-text field serves as a tool to keep track of the health of patients while also serving as a channel of communication between healthcare providers. It may even be argued that much of the clinically actionable information exists within the free text—contextual narrative of risk factors, symptoms, treatments, and assessment of quality of life. A recent study estimates average amount of free text per consultation was 68.2 words, compared to only 3.9 coded words, suggesting great value can be derived from the noncoded free-text field (Millares Martin, Reference Millares Martin2021). This may be even more true for mental health and older patients in managed care as ailment-specific details may be heterogenous. As such, Naismith etal.’s inclusion of free-text data adds a great deal of more information than that might be extracted from coded data alone (e.g., billing codes or diagnoses).

In understanding how NLP can support geropsychiatric research, it is important to understand its strengths and limitations. The free-text field (Wilton, Reference Wilton1991), as the name suggests, lacks the structure expected for automated processing. NLP is a subfield of artificial intelligence (AI) that strives to solve this problem. NLP has made considerable progress over the past decade including in healthcare. First, NLP systems are subject to a tradeoff—simplistic but transparent, or, sophisticated but difficult to assess the performance or perform replication (Naylor etal., Reference Naylor, French, Terker and Kamath2021). The study (Naismith etal., Reference Naismith, Howard, Stewart, Pitman and Mueller2022) derived associations between dementia subtypes and incidence of suicide ideation using NLP techniques on free-text fields. Using a relatively simple discoverable term matching method to identify records for suicide ideation while the diagnosis was identified based on coded fields, the study draws strength from both transparency and replicability. However, some clinical applications especially in mental health field may demand more complex models (Badal etal., Reference Badal2021a; Badal etal., Reference Badal2021b). Second, NLP system performance is not dependent not only on its own internal sophistication but also on the domain and the subject matter it is applied to. System performance depends upon language peculiarities such as possible overlap of terms with other diagnosis. To illustrate this point, a search on PubMed (White, Reference White2020) reveals “ideation” is almost always used in context of suicide in medical terminology. In contrast, “inflammation” has many more contexts: immunity, hypertension, irritable bowel syndrome (IBS), pain, gingival, and gonadal. This clearly suggests that the incidence of false positives, when associating diagnosis with inflammation using free-text field, will be much higher in such cases and the burden of disambiguation is unknowingly transferred to person interpreting the results.

From a clinical perspective, questioning of patient depends on symptoms, medical history, as well as on the interests and specialization of the healthcare provider. This has the potential to skew NLP-based findings. To illustrate this point, people with dementia with Lewy bodies (DLB) may experience recurrent visual hallucinations and psychosis symptoms (Morris etal., Reference Morris, Olichney and Corey-Bloom1998; Outeiro etal., Reference Outeiro2019), possibly causing increased questioning about suicide ideation. Further, DLB is often underdiagnosed or takes longer to diagnose (Jones and O'brien, Reference Jones and O'brien2014), which may result in a different presentation at evaluation compared to other dementia causes. The increased questioning associated with longer diagnosis times will usually contribute information to free-text fields, increasing the chances for false positives in NLP systems used for medical applications using health records. Further, the study uses a fixed search string “ideat*” to search for incidences of suicide ideation (Naismith etal., Reference Naismith, Howard, Stewart, Pitman and Mueller2022). The system may be enhanced to search for other terms “suicidal thoughts” in the free-text field.

Performance of NLP systems also depends upon the degree of adherence to controlled vocabulary in free-text field by medical workers who have contributed to it. Such vocabularies were conceived many years ago (Evans etal., Reference Evans, Cimino, Hersh, Huff and Bell1994) and are often defined and controlled by different parties. It would serve well for medical practitioners to familiarize themselves with some of these standards as they are adopted with varying speeds and degrees worldwide. Adhering to terminology defined by NCI Thesaurus (NCIt), for clinical care has had direct implications to basic and translational research as well as administrative activities. Systematized Nomenclature of Medicine Clinical Terms (SNOMED CT) was found to be extremely efficient in representing patient problems, and it had a sensitivity of 92.3%, a specificity of 80.0%, and a positive predictive value of 99.8% (Elkin etal., Reference Elkin2006). Logical Observation Identifiers Names and Codes (LOINC) (Uchegbu and Jing, Reference Uchegbu and Jing2017) offers a common language for clinical observations and laboratory tests. RxNorm (Nelson etal., Reference Nelson, Zeng, Kilbourne, Powell and Moore2011) has greatly facilitated normalization of generic and branded drug names and pharmacological compounds. It should be noted that such vocabularies are themselves evolving (Cimino and Clayton, Reference Cimino and Clayton1994).

Then there is a question how one translates NLP findings to clinical actions. With the convergence, standardization, and adoption of EHR, NLP is destined to play a crucial and even an everyday role in healthcare. Systems that recommend a course of action (e.g., coronary-artery bypass grafting surgery, CABG) (Lenert and Tovar, Reference Lenert and Tovar1993) and even prognosis (Jensen etal., Reference Jensen2017), based on individual records, have existed. Another paper in the recent issue of International Psychogeriatrics provides an example of contextual backdrop that promises the utility of such NLP-based systems. Not only depression, perceived social support, and attachment anxiety were found to be significantly associated with each other but also there were gender differences in levels of attachment anxiety playing a role in suicide ideation in the elderly (Oon-Arom etal., Reference Oon-Arom, Wongpakaran, Kuntawong and Wongpakaran2021). Some of this information may exist in coded fields (gender, age, and depression), while perceived social support and cues to attachment anxiety may be gleaned from historical EHR of patient expressed in free-text fields. NLP-enabled systems are uniquely capable of integrating these diverse pieces of information across the patient history and may even recommend interventions. The performance of such systems may, however, depend upon parameters beyond institutional control. Criticism for such uses stem from the modification of meaning that the free-text fields might bring to the otherwise objective information (Hogan and Wagner, Reference Hogan and Wagner1996).

Despite these words of caution about NLP findings, Naismith etal.’s work provides an excellent primer in how the clinical record can be used to support novel understanding on critical and emerging topics in geropsychiatry, such as suicide risk in newly diagnosed people with dementia. In the future, we would expect to see applications of NLP to other concerns in aging and mental health, which will derive a need for shared resources and understanding in quantifying the performance of such applications. Furthermore, findings derived from such systems must also be assessed on how they converge with that of more traditional research tools. The study by Naismith et. al. effectively addresses many of these points and provides a glimpse into the future of how NLP applied to clinical records can support understanding of emerging challenges in geropsychiatry.