Patients discharged with nonspecific diagnoses after acute hospital care: Characteristics, clustering and risk prediction

Nonspecific diagnoses after acute hospital care are registered if there is no established cause of a patient’s complaints when the patient is deemed fit for discharge. In this situation, time-critical conditions should have been ruled out. Nonspecific diagnoses account for 25–50% of all discharges after acute hospital care. Given the clinical context, studies have reported surprisingly high risks of mortality and readmission, possibly caused by missed diagnoses. Nonspecific diagnoses cover a variety of unresolved symptoms. The frequency and associated risk of these different symptoms remained to be determined. Identification of homogenous populations through unsupervised machine learning or individuals of increased risk through supervised machine learning could allow for interventions at discharge.

With this thesis and the four included manuscripts, we aimed to develop a method for forming hospital care episodes from individual DNPR contacts (Manuscript I), to investigate the frequency and associated risk of mortality and readmissions for nonspecific diagnoses and formed clinical subgroups of nonspecific diagnoses (Manuscript II), to identify homogenous clusters of patients with nonspecific diagnoses using unsupervised machine learning (Manuscript III), and with a basis in patients discharged with nonspecific abdominal pain, to develop a supervised machine model predicting 30-day readmissions and examine the causes of readmissions (Manuscript IV). All four studies were observational and register-based, using an established database at Statistics Denmark.

In Manuscript I, we developed an algorithm reading in and preparing DNPR data and forming hospital care episodes from sequential contacts. Based on sensitivity analysis and visualization, we suggested 4 hours as a window for combining individual contacts. The algorithm was developed in collaboration with a team of experts and has been made freely and publicly available at the online repository Figshare.

In Manuscript II, we identified a cohort of 192,185 patients discharged after hospital care episodes initiated in emergency departments. Nonspecific discharge diagnoses were registered for 51% of short hospital care episodes (3–12 hours) and 26% of lengthier episodes (12–168 hours). Nonspecific diagnoses had a lower absolute risk of both mortality and readmissions than disease-specific diagnoses. However, the relative risk was similar for mortality and minor for readmissions after adjusting for relevant confounders. From the registered discharge diagnoses, we identified 50 different clinical subgroups (such as chest pain, vertigo, or malaise and fatigue), varying notably in size, average characteristics and prognosis after discharge.

In Manuscript III, we conducted mixed-data unsupervised machine learning on a cohort of 92,650 patients discharged with nonspecific diagnoses. After manual revision with a team of experts, we chose a k-prototypes model identifying 20 clusters as the one best describing distinct clusters. The clusters varied in size and prognosis and described clinically recognizable subgroups of patients.

In Manuscript IV, we developed a decision tree predicting readmissions on a cohort of 13,445 patients discharged with nonspecific abdominal pain. Readmissions occurred for 10.6%. The decision tree utilizing generic information such as age, comorbidity score and proportion of abnormal biochemical results, achieved a sensitivity of 17% and positive predictive value of 24% when predicting readmission for 8% of the population. The model performed poorly on two clusters with low and average mean age—but models developed specifically on these clusters were also found insufficient. We further reported that approximately 50% of readmissions are likely caused by missed diagnoses, whereas the other 50% are either due to anxiety from ongoing symptoms or unrelated causes.

With this thesis we conclude that nonspecific diagnoses pose a healthcare problem. The adverse outcomes are likely in part due to missed diagnoses, unresolved anxiety, and unrelated causes, which should all be sought reduced going forward.