Fairness in Predicting Cancer Mortality Across Racial Subgroups
1. A machine learning model trained on retrospective data of patients with solid malignant tumours was able to predict mortality with similar performance for all racial subgroups, indicating no evidence of racial bias.
Evidence Rating Level: 2 (Good)
While machine learning (ML) models are powerful tools that have the potential to assist physicians in clinical decision-making, an important barrier to their implementation is the possibility for racial bias to exacerbate existing healthcare inequities. This cohort study therefore sought to investigate whether an ML model can accurately predict cancer mortality with similar performance across different racial subgroups. A dataset was obtained by extracting data from patients in the Mount Sinai Health System (MSHS) cancer registry with a cancer diagnosis between January 2016 and December 2021. 70% of the dataset was used to train the model using a random forest algorithm while the remaining data was used as a test dataset consisting of data from 43,274 patients (mean [SD] age = 64.09 [14.26] years). The study’s primary outcome was discriminatory performance, assessed using the F1 score and the area under the receiver operating characteristic (AUROC) curve, and fairness metrics for each racial subgroup (Asian, Black, Native American, White and other). Across all patients, the model performed with an F1 score of 0.35 (95% CI, 0.35-0.35) and AUROC of 0.76 (95%, 0.75-0.77). Performance was similar across all racial subgroups, with an AUROC of 0.75 (95% CI, 0.72-0.78) for Asian patients, 0.75 (95% CI, 0.73-0.77) for Black patients and 0.77 (95% CI, 0.75-0.79) for patients classified as other. Fairness metrics were also within pre-specified thresholds, indicating equal opportunity, equalized odds and no disparate impact using comparisons between racial subgroups. Overall, this study found no evidence for racial bias in the ML model’s performance, providing support for the role of assessing racial bias as an important component of ML model development.
Internet-Guided Cognitive Behavioral Therapy for Insomnia Among Patients With Traumatic Brain Injury
1. Patients with insomnia following traumatic brain injury randomized to complete an automated 6-week internet-guided cognitive behavioural therapy program were observed to have a greater improvement of their insomnia compared to sleep education.
2. Among patients completing an internet-guided cognitive behavioural therapy program, greater improvements in insomnia were correlated with greater improvement in depression, post-traumatic stress disorder, sleep quality and fatigue.
Evidence Rating Level: 1 (Excellent)
Individuals suffering from traumatic brain injury (TBI) frequently report suffering from insomnia. While cognitive behavioural therapy for insomnia (CBT-I) has been established as the first-line treatment for insomnia, CBT-I carries many barriers to implementation including a shortage of clinicians trained in CBT-I and financial constraints. To address these barriers, an automated internet-based adaptation of CBT-I (eCBT-I) has previously been introduced but has never been validated in patients with TBI or military service members. This randomized clinical trial therefore sought to investigate the efficacy of eCBT-I in US military service members with TBI and moderate to severe insomnia. 106 participants (mean [SD] age, 42 [12] years) were randomized in a 3:1 ratio to eCBT-I or a control intervention involving internet-based sleep education. Participants were assessed at baseline, at 9 weeks post-intervention and at a 3-month follow-up with the primary outcome measure being the change in insomnia severity index (ISI) score between baseline and post-intervention assessment. The reduction in ISI score for those randomized to eCBT-I was statistically significant at -6.0 compared to -2.3 for those randomized to sleep education, with a difference in the extent of improvement in ISI score of -3.5 (95% CI, −6.5 to −0.4; Cohen d, −0.32 [95% CI, −0.70 to −0.04]; P = .03). For those randomized to eCBT-I, there was a correlation between the extent of improvement in insomnia and the extent of improvement in depression severity at the post-intervention assessment (Spearman ρ = 0.68 [P < .001]) and the 3-month follow-up (Spearman ρ = 0.56 [P = .001]). Overall, this study found that a remote, internet-based alternative to cognitive behavioural therapy for insomnia was effective in improving insomnia symptoms in military service members with TBI
1. The use of metformin in patients with inflammatory bowel disease and type 2 diabetes mellitus was associated with improved clinical outcomes, such as a decreased need for steroids and IBD-related surgeries throughout a 3-year follow-up period.
Evidence Rating Level: 2 (Good)
Metformin is a first-line therapy used in the treatment of type 2 diabetes mellitus (T2DM) while also possessing several anti-inflammatory properties. While several studies in animal models of colitis have demonstrated metformin’s ability to alter inflammatory signalling, there are no studies evaluating the potential role of metformin as a therapeutic for IBD. This retrospective cohort study therefore sought to evaluate the outcomes of IBD patients on metformin for T2DM compared with matched controls of IBD patients with T2DM not receiving metformin. 1323 patients with ulcerative colitis (UC) (mean age = 59.5 ± 12.2 years) and 1278 patients with Crohn’s disease (CD) (mean age = 56.3 ± 12.6 years) were identified using the TriNetX database to form a metformin cohort. At 1 year following initiation of metformin, UC patients in the metformin cohort were at a lower risk for oral steroids (aOR, 0.61; 95% CI, 0.5-0.73; P < .01) and IV steroids (aOR, 0.45; 95% CI, 0.34-0.59; P < .01) compared to the control cohort. At 2 and 3 years, UC patients in the metformin cohort were at a lower risk for oral and IV steroids compared to the control cohort. Similarly, CD patients in the metformin cohort exhibited lower risk for oral and IV steroids compared to the control cohort at 1, 2 and 3 years of metformin use. There was no significant difference in the risk for IBD-related surgeries between UC patients in either cohort. However, CD patients in the metformin cohort were at a lower risk for IBD-related surgeries at 1 year (aOR, 0.5; 95% CI, 0.31-0.81; P < .01), 2 years (aOR, 0.54; 95% CI, 0.36-0.81; P < .01) and 3 years (aOR, 0.62; 95% CI, 0.43-0.89; P < .01) compared to the control cohort. Overall, this study found that the use of metformin in patients with IBD and T2DM was associated with improved outcomes such as a reduced need for oral and IV steroids, and supports further investigation into the possible role of metformin in IBD management.
1. Induction of a ketogenic diet was found to be safe and feasible for critically ill patients with sepsis, with no adverse effects observed on general lab parameters.
2. Gene expression of markers of T cell activation and several inflammatory markers were reduced in patients under a ketogenic diet, revealing a potential role for ketogenic diets in improving immune dysregulation in sepsis patients.
Evidence Rating Level: 1 (Excellent)
While carbohydrates are the recommended primary energetic metabolite for critically ill adult patients, they may trigger abnormal activation of immune cells and can therefore interfere with immune resolution in conditions involving excessive immune activation such as sepsis. This randomized controlled trial therefore sought to investigate whether induction of a stable ketotic state in critically ill adult patients with sepsis using a ketogenic diet (KD) is feasible. Between January 2020 and January 2022, 40 patients (mean age = 67.0 ± 13.1 years) with sepsis being treated in the intensive-care unit (ICU) of the University Hospital Knappschaftskrankenhaus Bochum were randomized to a KD group (n = 20) or a control group (n = 20). Participants in the KD group received a ketogenic formulation while those in the control group received a standard enteral nutrition solution. The primary study outcome was serum β-hydroxybutyrate (BHB) concentration on day 14 while secondary outcomes included feasibility, safety and outcome measures. On day 14, participants in the KD group experienced a significantly greater increase in BHB compared to patients on standard nutrition with an estimated mean difference of 1.4mM (95% CI: 1.0 to 1.8; P < 0.001) between groups. Additionally, participants in the KD group also experienced increased ventilation-free, vasopressor-free, dialysis-free and ICU-free days compared to controls. At day 14, the KD group experienced significant differential expression of genes involved in T cell activation and signalling compared to the control group. Overall, this study found that induction of a stable state of ketosis is feasible in critically ill adult patients with sepsis on a ketogenic diet, and that a ketogenic diet may carry therapeutic potential in sepsis by addressing immune dysregulation.
1. When presented with thyroid ultrasounds, an artificial intelligence system developed using deep learning exhibited a diagnostic performance comparable to senior radiologists.
Evidence Rating Level: 2 (Good)
Detection of thyroid nodules is extremely common using ultrasound, but differentiation of these nodules into benign or malignant cases is highly variable and dependent on the level of experience of the reviewing radiologist. This multicentre prospective study therefore sought to investigate whether an artificial intelligence (AI) system could assist in clinical decision-making for junior radiologists when reviewing ultrasounds of thyroid nodules. The AI system, AI-SONIC™ Thyroid, was presented with ultrasound scans of thyroid nodules from 1036 patients from three medical centres (University of Chinese Academy of Sciences, the Second Affiliated Hospital of Shantou University and the Second Affiliated Hospital of Nanchang University) and asked to make diagnostic predictions based on an established two-level hierarchical system (2e diagnostic criteria). Two junior and two senior radiologists made independent classifications of pro-benign or pro-malignant, and six months later repeated their diagnoses with assistance from the AI system. The AI system had comparable sensitivity, specificity and accuracy compared to senior radiologists (0.868 vs. 0.849, 0.934 vs. 0.934, 0.917 vs. 0.912; all p > 0.05) while showing superior sensitivity, specificity and accuracy compared to junior radiologists (0.868 vs. 0.820, 0.934 vs. 0.837, 0.917 vs. 0.833; all p < 0.001). AI-assisted readings improved the specificity and accuracy of junior radiologists compared to independent readings (0.837 vs. 0.921, 0.833 vs. 0.895; all p < 0.001) but did not improve the specificity and accuracy of senior radiologists (all p > 0.05). Overall, this study found that an AI system is capable of achieving a diagnostic performance on thyroid ultrasounds comparable to senior radiologists and may be an important tool to enhance the diagnostic performance of junior radiologists.
Image: PD
©2024 2 Minute Medicine, Inc. All rights reserved. No works may be reproduced without expressed written consent from 2 Minute Medicine, Inc. Inquire about licensing here. No article should be construed as medical advice and is not intended as such by the authors or by 2 Minute Medicine, Inc.