Development and validation of 'Patient Optimizer' (POP) algorithms for predicting surgical risk with machine learning.

BACKGROUND: Pre-operative risk assessment can help clinicians prepare patients for surgery, reducing the risk of perioperative complications, length of hospital stay, readmission and mortality. Further, it can facilitate collaborative decision-making and operational planning. OBJECTIVE: To develop effective pre-operative risk assessment algorithms (referred to as Patient Optimizer or POP) using Machine Learning (ML) that predict the development of post-operative complications and provide pilot data to inform the design of a larger prospective study. METHODS: After institutional ethics approval, we developed a base model that encapsulates the standard manual approach of combining patient-risk and procedure-risk. In an automated process, additional variables were included and tested with 10-fold cross-validation, and the best performing features were selected. The models were evaluated and confidence intervals calculated using bootstrapping. Clinical expertise was used to restrict the cardinality of categorical variables (e.g. pathology results) by including the most clinically relevant values. The models were created with logistic regression (LR) and extreme gradient-boosted trees using XGBoost (Chen and Guestrin, 2016). We evaluated performance using the area under the receiver operating characteristic curve (AUROC) and the area under the precision-recall curve (AUPRC). Data was obtained from a metropolitan university teaching hospital from January 2015 to July 2020. Data collection was restricted to adult patients undergoing elective surgery. RESULTS: A total of 11,475 adult admissions were included. The performance of XGBoost and LR was very similar across endpoints and metrics. For predicting the risk of any post-operative complication, kidney failure and length-of-stay (LOS), POP with XGBoost achieved an AUROC (95%CI) of 0.755 (0.744, 0.767), 0.869 (0.846, 0.891) and 0.841 (0.833, 0.847) respectively and AUPRC of 0.651 (0.632, 0.669), 0.336 (0.282, 0.390) and 0.741 (0.729, 0.753) respectively. For 30-day readmission and in-patient mortality, POP with XGBoost achieved an AUROC (95%CI) of 0.610 (0.587, 0.635) and 0.866 (0.777, 0.943) respectively and AUPRC of 0.116 (0.104, 0.132) and 0.031 (0.015, 0.072) respectively. CONCLUSION: The POP algorithms effectively predicted any post-operative complication, kidney failure and LOS in the sample population. A larger study is justified to improve the algorithm to better predict complications and length of hospital stay. A larger dataset may also improve the prediction of additional specific complications, readmission and mortality.

Natural language processing diagnosed behavioural disturbance phenotypes in the intensive care unit: characteristics, prevalence, trajectory, treatment, and outcomes.

BACKGROUND: Natural language processing (NLP) may help evaluate the characteristics, prevalence, trajectory, treatment, and outcomes of behavioural disturbance phenotypes in critically ill patients. METHODS: We obtained electronic clinical notes, demographic information, outcomes, and treatment data from three medical-surgical ICUs. Using NLP, we screened for behavioural disturbance phenotypes based on words suggestive of an agitated state, a non-agitated state, or a combination of both. RESULTS: We studied 2931 patients. Of these, 225 (7.7%) were NLP-Dx-BD positive for the agitated phenotype, 544 (18.6%) for the non-agitated phenotype and 667 (22.7%) for the combined phenotype. Patients with these phenotypes carried multiple clinical baseline differences. On time-dependent multivariable analysis to compensate for immortal time bias and after adjustment for key outcome predictors, agitated phenotype patients were more likely to receive antipsychotic medications (odds ratio [OR] 1.84, 1.35-2.51, p < 0.001) compared to non-agitated phenotype patients but not compared to combined phenotype patients (OR 1.27, 0.86-1.89, p = 0.229). Moreover, agitated phenotype patients were more likely to die than other phenotypes patients (OR 1.57, 1.10-2.25, p = 0.012 vs non-agitated phenotype; OR 4.61, 2.14-9.90, p < 0.001 vs. combined phenotype). This association was strongest in patients receiving mechanical ventilation when compared with the combined phenotype (OR 7.03, 2.07-23.79, p = 0.002). A similar increased risk was also seen for patients with the non-agitated phenotype compared with the combined phenotype (OR 6.10, 1.80-20.64, p = 0.004). CONCLUSIONS: NLP-Dx-BD screening enabled identification of three behavioural disturbance phenotypes with different characteristics, prevalence, trajectory, treatment, and outcome. Such phenotype identification appears relevant to prognostication and trial design.

Decompensated metabolic acidosis in the emergency department: Epidemiology, sodium bicarbonate therapy, and clinical outcomes.

Objective: This article aims to describe the epidemiology of decompensated metabolic acidosis, the characteristics of sodium bicarbonate (SB) administration and outcomes in emergency department (ED) patients. Design: This is a retrospective cohort study. Setting: ED of a tertiary referral hospital in Melbourne, Australia. Participants: Adult patients presenting to the ED between 1 July 2011 and 20 September 2020 with decompensated metabolic acidosis diagnosed on arterial blood gas (ABG). Main outcome measures: We compared characteristics between those treated with or without SB. We studied SB administration characteristics, change in laboratory variables, factors associated with use and dose, and clinical outcomes. Results: Among 753,613 ED patients, 314 had decompensated metabolic acidosis on ABG, with 17.8% receiving SB. Patients in the SB group had lower median pH, CO2, bicarbonate, and base excess (BE) levels compared with the No SB group (P < 0.01). The median number of SB doses in the SB group was one treatment. This was given at a median total dose of 100 mmol and at a median of 2.8 h after the diagnostic blood gas results. Only 42% of patients in the SB group had a subsequent blood gas measured. In such patients, there was no significant change in pH, bicarbonate, or BE. SB therapy was not independently associated with mortality. Conclusions: ABG-confirmed decompensated metabolic acidosis was rare but associated with a high mortality. SB administration occurred in a minority of patients and in more acidaemic patients. However, SB dose was stereotypical and not tailored to acidosis severity. Assessment of SB effect was infrequent and showed no correction of acidosis. Systematic studies of titrated SB therapy are required to inform current practice.

The epidemiology of ketosis and low bicarbonate concentration in inpatients treated with sodium-glucose linked cotransporter inhibitors or dipeptidyl peptidase-4 inhibitors.

AIMS: To compare the level of ketones and bicarbonate in inpatients treated with sodium-glucose linked cotransporter 2 inhibitors (SGLT2i) and those treated with dipeptidyl peptidase-4 inhibitors (DPP4i). METHODS: We conducted an electronic medical records-based cohort study. We identified patients with type 2 diabetes, with ketone measurements available, who received SGLT2i (n = 82) or DPP4i (n = 308) during admission. We compared ketone levels between those who received SGLT2i or DPP4i using mixed ordinal logistic regression. The primary outcome was level of ketosis (<0.6, 0.6-1.5, 1.6-3.0, >3 mmol/L). Secondary outcomes included bicarbonate levels, hospital complications, ICU admission, and death. RESULTS: SGLT2i use was not associated with greater ketosis than DPP4i use, after adjusting for age, weight, Charlson Comorbidity Index, HbA1c, estimated glomerular filtration rate, principal diagnosis category, admission type and insulin administration (OR 4.52 95 % CI (0.33, 61.82)). After adjustment, there was no difference in complications (p = 0.14), ICU admissions (p = 0.64), mortality (p = 0.30), or bicarbonate levels (p = 0.97). CONCLUSION: Ketone levels were not greater in patients who received SGLT2i than those who received DPP4i. There were no differences in bicarbonate levels, complications, ICU admissions, or mortality, implying that, in inpatients, SGLT2i use is neither associated with ketosis nor adverse clinical outcomes.

Early and prolonged continuous hypertonic saline infusion in patients with acute liver failure.

PURPOSE: To study patient characteristics, physiological changes, and outcomes associated with prolonged continuous hypertonic saline (HTS) infusion in acute liver failure (ALF). MATERIALS AND METHODS: Retrospective observational cohort study of adult patients with ALF. We collected clinical, biochemical, and physiological data six hourly for the first week, daily until day 30 or hospital discharge, and weekly, when documented, until day 180. RESULTS: Of 127 patients, 85 received continuous HTS. Compared with non-HTS patients they were more likely to receive continuous renal replacement therapy (CRRT) (p < 0.001) and mechanical ventilation (p < 0.001). Median HTS duration was 150 (Interquartile range (IQR): 84-168) hours, delivering a median 2244 (IQR: 979-4610) mmol sodium load. Median peak sodium concentration was 149 mmol/L vs 138 mmol/L in non-HTS patients (p < 0.001). The median rate of sodium increase with infusion was 0.1 mmol/L/h and median rate of decrease during weaning was 0.1 mmol/L every 6 h. Median lowest pH value was 7.29 vs. 7.35 in non-HTS patients. Survival of HTS patients was 72.9% overall and 72.2% without transplantation. CONCLUSIONS: In ALF patients, the prolonged administration of HTS infusion was not associated with severe hypernatremia or rapid shifts in serum sodium upon commencement, delivery, or weaning.

Early prediction of hospital admission of emergency department patients.

OBJECTIVE: The early prediction of hospital admission is important to ED patient management. Using available electronic data, we aimed to develop a predictive model for hospital admission. METHODS: We analysed all presentations to the ED of a tertiary referral centre over 7 years. To our knowledge, our data set of nearly 600 000 presentations is the largest reported. Using demographic, clinical, socioeconomic, triage, vital signs, pathology data and keywords in electronic notes, we trained a machine learning (ML) model with presentations from 2015 to 2020 and evaluated it on a held-out data set from 2021 to mid-2022. We assessed electronic medical records (EMRs) data at patient arrival (baseline), 30, 60, 120 and 240 min after ED presentation. RESULTS: The training data set included 424 354 data points and the validation data set 53 403. We developed and trained a binary classifier to predict inpatient admission. On a held-out test data set of 121 258 data points, we predicted admission with 86% accuracy within 30 min of ED presentation with 94% discrimination. All models for different time points from ED presentation produced an area under the receiver operating characteristic curve (AUC) ≥0.93 for admission overall, with sensitivity/specificity/F1-scores of 0.83/0.90/0.84 for any inpatient admission at 30 min after presentation and 0.81/0.92/0.84 at baseline. The models retained lower but still high AUC levels when separated for short stay units or inpatient admissions. CONCLUSION: We combined available electronic data and ML technology to achieve excellent predictive performance for subsequent hospital admission. Such prediction may assist with patient flow.

Breaches of pre-medical emergency team call criteria in an Australian hospital.

Objectives and outcomes: To evaluate the 24hrs before medical emergency team (MET) calls to examine: 1) the frequency, nature, and timing of pre-MET criteria breaches; 2) differences in characteristics and outcomes between patients who did and didn't experience pre-MET breaches. Design: Retrospective observational study November 2020-June 2021. Setting: Tertiary referral Australian hospital. Participants: Adults (≥18 years) experiencing MET calls. Results: Breaches in pre-MET criteria occurred prior to 1886/2255 (83.6%) MET calls, and 1038/1281 (81.0%) of the first MET calls. Patients with pre-MET breaches were older (median [IQR] 72 [57-81] vs 66 [56-77] yrs), more likely to be admitted from home (87.8% vs 81.9%) and via the emergency department (73.0% vs 50.2%), but less likely to be for full resuscitation after (67.3% vs 76.5%) the MET. The three most common pre-MET breaches were low SpO2 (48.0%), high pulse rate (39.8%), and low systolic blood pressure (29.0%) which were present for a median (IQR) of 15.4 (7.5-20.8), 13.2 (4.3-21.0), and 12.6 (3.5-20.1) hrs before the MET call, respectively. Patients with pre-MET breaches were more likely to need intensive care admission within 24 h (15.6 vs 11.9%), have repeat MET calls (33.3 vs 24.7%), and die in hospital (15.8 vs 9.9%). Conclusions: Four-fifths of MET calls were preceded by pre-MET criteria breaches, which were present for many hours. Such patients were older, had more limits of treatment, and experienced worse outcomes. There is a need to improve goals of care documentation and pre-MET management of clinical deterioration.

Prevalence, severity, duration and resolution of cholestasis after acute liver failure.

OBJECTIVE: Persistent cholestasis may follow acute liver failure (ALF), but its course remains unknown. We aimed to describe the prevalence, onset, severity, duration and resolution of post-ALF cholestasis. DESIGN: Cohort of 127 adult patients with ALF at a liver transplantation centre identified using electronic databases. We obtained laboratory data every 6 hours for the first week, daily until day 30 and weekly, when documented, until day 180. RESULTS: Median age was 40.7 (IQR 31.0-52.4) years, median peak alanine aminotransferase level was 5494 (2521-8819) U/L and 87 (68.5%) cases had paracetamol toxicity. Overall, 12.6% underwent transplantation (3.4% for paracetamol vs 32.5% for non-paracetamol; p<0.001). Ninety-day mortality was 20.7% for paracetamol versus 30.0% for non-paracetamol patients. All non-transplanted survivors reached a bilirubin level>50 µmol/L, which peaked 3.5 (1.0-10.1) days after admission at 169.0 (80.0-302.0) µmol/L. At hospital discharge, 18.8% of patients had normal bilirubin levels and, at a median follow-up time from admission to last measurement of 16 (10-30) days, 46.9% had normal levels. Similarly, there was an increase in alkaline phosphatase (ALP) (207.0 (148.0-292.5) U/L) and gamma-glutamyl transferase (GGT) (336.0 (209.5-554.5) U/L) peaking at 4.5 days, with normalised values in 40.3% and 8.3% at hospital discharge. CONCLUSION: Post-ALF cholestasis is ubiquitous. Bilirubin, ALP and GGT peak at 3 to 5 days and, return to baseline in the minority of patients at median follow-up of 16 days. These data inform clinical expectations of the natural course of this condition.

Natural language processing diagnosed behavioral disturbance vs confusion assessment method for the intensive care unit: prevalence, patient characteristics, overlap, and association with treatment and outcome.

PURPOSE: To compare the prevalence, characteristics, drug treatment for delirium, and outcomes of patients with Natural Language Processing (NLP) diagnosed behavioral disturbance (NLP-Dx-BD) vs Confusion Assessment Method for intensive care unit (CAM-ICU) positivity. METHODS: In three combined medical-surgical ICUs, we obtained data on demographics, treatment with antipsychotic medications, and outcomes. We applied NLP to caregiver progress notes to diagnose behavioral disturbance and analyzed simultaneous CAM-ICU. RESULTS: We assessed 2313 patients with a median lowest Richmond Agitation-Sedation Scale (RASS) score of - 2 (- 4.0 to - 1.0) and median highest RASS score of 1 (0 to 1). Overall, 1246 (53.9%) patients were NLP-Dx-BD positive (NLP-Dx-BDpos) and 578 (25%) were CAM-ICU positive (CAM-ICUpos). Among NLP-Dx-BDpos patients, 539 (43.3%) were also CAM-ICUpos. In contrast, among CAM-ICUpos patients, 539 (93.3%) were also NLP-Dx-BDpos. The use of antipsychotic medications was highest in patients in the CAM-ICUpos and NLP-Dx-BDpos group (24.3%) followed by the CAM-ICUneg and NLP-Dx-BDpos group (10.5%). In NLP-Dx-BDneg patients, antipsychotic medication use was lower at 5.1% for CAM-ICUpos and NLP-Dx-BDneg patients and 2.3% for CAM-ICUneg and NLP-Dx-BDneg patients (overall P < 0.001). Regardless of CAM-ICU status, after adjustment and on time-dependent Cox modelling, NLP-Dx-BD was associated with greater antipsychotic medication use. Finally, regardless of CAM-ICU status, NLP-Dx-BDpos patients had longer duration of ICU and hospital stay and greater hospital mortality (all P < 0.001). CONCLUSION: More patients were NLP-Dx-BD positive than CAM-ICU positive. NLP-Dx-BD and CAM-ICU assessment describe partly overlapping populations. However, NLP-Dx-BD identifies more patients likely to receive antipsychotic medications. In the absence of NLP-Dx-BD, treatment with antipsychotic medications is rare.

The comparative epidemiology and outcomes of hospitalized patients treated with SGLT2 or DPP4 inhibitors.

OBJECTIVE: To compare the outcomes of sodium glucose linked cotransporter 2 inhibitors (SGLT2i) and dipeptidyl peptidase 4 inhibitors (DPP4i) in hospitalized patients. RESEARCH DESIGN AND METHODS: Electronic medical records-based cohort study. Identification of patients with type 2 diabetes and treatment with SGLT2i (n = 466) or DPP4i (n = 1541). Outcomes compared between those who received SGLT2i and those who received DPP4i. The primary outcome: adjusted percentage of blood glycemia within 4-10 mmol/L. RESULTS: After adjustment, SGLT2i use had a statistically equivalent percentage of glycemia within range (coefficient: 4.55, 95% CI -3.23 to 12.32, p = 0.25) or <4 mmol/L (coefficient -0.17, 95% CI -0.71 to 3.72, p = 0.54). There were no significant differences in hospital length of stay (p = 0.22), complications, (p = 0.11) or mortality (p = 0.57). When measured, ketone levels were higher in the SGLT2i group on admission, but lower on days 3, 4 and 5 (p < 0.001 for interaction). Bicarbonate levels were not statistically different between groups. Finally, 54% of patients whose SGLT2i was ceased during admission, were discharged home without it. CONCLUSION: Among inpatients with type 2 diabetes, SGLT2i use was associated with equivalent within-target glycaemia and no significant increase in hypoglycemia, ketonemia, or lower bicarbonate levels. These hypothesis-generating findings support further investigation of SGLT2i therapy in inpatients.

Performance of 4 Creatinine-based Equations in Assessing Glomerular Filtration Rate in Adults with Diabetes.

AIMS: To evaluate diagnostic performance of glomerular filtration rate (GFR) estimated by modification of diet in renal disease (MDRD), chronic kidney disease epidemiology collaboration (CKD-EPI), full age spectrum (FAS), and revised Lund-Malmö (r-LM) equations in adults with diabetes. METHODS: Individuals were included in this cross-sectional study if they had at least 1 measurement of technetium-99m diethylenetriamine-pentaacetic acid (99mTc-DTPA) GFR (mGFR) and serum creatinine (1487 patients with 2703 measures). GFR calculated by estimation equations was compared with mGFR. Diagnostic performance was assessed using concordance correlation coefficient (CCC), bias, precision, accuracy, reduced major axis regression (RMAR), and Bland-Altman plot. Analysis was repeated in subgroups based on sex, diabetes type, Hemoglobin A1C, and GFR level. RESULTS: Of all patients, 1189 (86%) had type 2 diabetes. Mean mGFR, MDRD, CKD-EPI, FAS, and revised Lund-Malmö eGFR were 66, 72, 74, 71, and 67 mL/min/1.73m2, respectively. Overall, the r-LM had the highest CCC (0.83), lowest bias (-1.4 mL/min/1.73 m2), highest precision (16.2 mL/min/1.73 m2), and highest accuracy (P10 = 39%). The RMAR (slope, intercept) in r-LM, FAS, MDRD, and CKD-EPI was 1.18, -13.35; 0.97, -2.9; 1, -6.4, and 1.04, -11.3, respectively. The Bland-Altman plot showed that r-LM had the lowest mean difference and the narrowest 95% limit of agreement (-1.0, 54.1 mL/min/1.73 m2), while mean difference was more than 5-fold higher in FAS, MDRD, and CKD-EPI (-5.2, -6.3, and -8.2, respectively). CONCLUSIONS: In adults with diabetes the revised Lund-Malmö performs better than MDRD, CKD-EPI, and FAS in calculating point estimates of GFR.