28 March 2025 in Science Advances
Advances in artificial intelligence (AI) have achieved expert-level performance in medical imaging applications. Notably, self-supervised vision-language foundation models can detect a broad spectrum of pathologies without relying on explicit training annotations. However, it is crucial to ensure that these AI models do not mirror or amplify human biases, disadvantaging historically marginalized groups such as females or Black patients. In this study, we investigate the algorithmic fairness of state-of-the-art vision-language foundation models in chest x-ray diagnosis across five globally sourced datasets. Our findings reveal that compared to board-certified radiologists, these foundation models consistently underdiagnose marginalized groups, with even higher rates seen in intersectional subgroups such as Black female patients. Such biases present over a wide range of pathologies and demographic attributes. Further analysis of the model embedding uncovers its substantial encoding of demographic information. Deploying medical AI systems with biases can intensify preexisting care disparities, posing potential challenges to equitable healthcare access and raising ethical questions about their clinical applications.
1 April 2018 in Antimicrobial Agents and Chemotherapy
We performed a multicenter, prospective, randomized study to investigate the efficacy and safety of clofazimine (CLO) for treatment of extensively drug-resistant tuberculosis (XDR-TB) in China. Forty-nine patients infected with XDR-TB were randomly assigned to either the control group or the CLO group, both of which received 36 months of individually customized treatment. The primary endpoint was the time to sputum culture conversion on solid medium. Clinical outcomes of patients were evaluated at the time of treatment completion. Of the 22 patients in the experimental group, 7 (31.8%) met the treatment criterion of “cure” and 1 (4.5%) “complete treatment,” for a total of 8 (36.4%) exhibiting successful treatment outcomes without relapse. In the control group, 6 patients (22.2%) were cured and 6 (22.2%) completed treatment by the end of the study. Statistical analysis revealed no significant difference in successful outcome rates between the CLO group and the control group. The average sputum culture conversion time for the experimental group was 19.7 months, which was not statistically different from that for the control group (20.3 months; P = 0.57). Of the 22 patients in the CLO group, 12 (54.5%) experienced adverse events after starting CLO treatment. The most frequently observed adverse event was liver damage, with 31.8% of patients (7/22 patients) in the CLO group versus 11.1% (3/27 patients) in the control group exhibiting this adverse event. Our study demonstrates that inclusion of CLO in background treatment regimens for XDR-TB is of limited benefit, especially since hepatic disorders arise as major adverse events with CLO treatment. (This study is registered with the Chinese Clinical Trial Registry [ChiCTR, www.chictr.org.cn ] under identifier ChiCTR1800014800.)
14 February 2023 in European Radiology
SPSudhakar N. J. Pipavath
+1
25 December 2024 in CHEST Pulmonary
Pulmonary nodules (PNs) are frequently detected by chest CT scan, which is increasingly used in clinical practice. Accurately identifying malignant nodules can pose a diagnostic challenge; therefore, a high-specificity biomarker could help clinicians identify malignant nodules and ideally lead to the earlier diagnosis of lung cancer. What are the performance characteristics of a blood-based biomarker for identifying malignancy in patients with a CT-detected PN? Banked plasma samples from 2 independent prospective observational cohorts of patients presenting with benign or malignant PNs 8 to 30 mm in size were tested using a 7-autoantibody panel. Sensitivity, specificity, and positive predictive value of the autoantibody test (AAT) to identify cancer were calculated for the individual and combined cohorts. Overall, 447 patients (263 and 184 from each cohort) were included in the analysis with a prevalence of malignancy of 55%. The performance of the AAT between the 2 cohorts was similar. The AAT demonstrated a specificity of 90% (95% CI, 85%-93%), a positive predictive value of 66% (95% CI, 52%-77%), sensitivity of 16% (95% CI, 12%-22%), and false-positive rate of 10% in the combined cohort. Using a pretest probability of cancer cutoff of 20% improved the positive predictive value to 76% (95% CI, 61%-88%) and resulted in a 52% decrease in the number of false-positive test results. In the subset of patients who had 18F-fluorodeoxyglucose PET imaging performed for clinical purposes (n = 222), specificity of the AAT was higher (93% vs 58%, P < .001), but the sensitivity was lower than 18F-fluorodeoxyglucose PET scan (17% vs 75%, P < .001). This study validates the specificity of a blood-based autoantibody biomarker for identifying malignancy in patients with indeterminate PNs. This rule-in biomarker may help to expedite workup of malignant nodules. ClinicalTrials.gov; No.: NCT01752114; URL: www.clinicaltrials.gov CHEST Pulmonary 2025; 3(1):100130
1 July 2025 in Cancer Immunology Research
The blood T-cell receptor (TCR) repertoire broadly reflects current and lifetime immune responses against infectious pathogens and cancer, but the circulating T-cell repertoire remains a largely untapped resource for cancer biomarker studies due to repertoire complexity and limited profiling data. In this study, we investigated the use of blood TCR sequencing for the early detection of lung cancer. We sequenced the leukocyte fraction of peripheral blood from 633 individuals divided into a case–control cohort (n = 511) and a lung cancer screening cohort (n = 122), representing more than 12.6 million unique clonotypes. Based on the TCR repertoires in these individuals, we devised a Tumor Immune Lymphocyte Score (TILS) using either TCR specificity groups (TILS-A) or highly recurrent “public” TCR clonotypes (TILS-B) capable of detecting lung cancer. TILS-A consisted of 125 TCR specificity groups that outperformed the TILS-B classifier of 49 public, TCRβ-Vβ–defined clonotypes for cancer detection. TILS classifiers (TILS-A and TILS-B) provided predictive value after accounting for age, smoking status, and nodule size in the lung cancer screening cohort and improved cancer prediction for individuals with indeterminate lung cancer risk. In the subgroup analysis, TILS-A was associated with lung cancer in both early- and late-stage disease, had improved accuracy when accounting for HLA status, and was validated in an external dataset studying lung cancer initiation. Collectively, these data suggest that profiles of the circulating T-cell response can provide value for lung cancer detection and support its use as a diagnostic tool.