Machine Learning Model Outperforms Pulmonologists in Detecting Lung Cancer, suggests research
Researchers have discovered that a machine learning (ML) model, using dynamic ensemble selection (DES), outperforms experienced pulmonologists in cancer lung diagnosis. Lung cancer is the cause of cancer deaths in the world, mainly due to late detection. Improving survival rates is dependent on the early detection approach. A recent study was conducted by Ricco N. and colleagues which was published in the journal of Scientific Reports .
This was a retrospective analysis of data from 38,944 patients suspected to have LC within the Region of Southern Denmark from 2009 to 2018. The study included 9,940 patients with complete data, with 2,505 (25%) diagnosed with LC. The DES model involved smoking history and key blood biomarkers such as lactate dehydrogenase, total calcium, sodium levels, leukocyte and neutrophil counts, and C-reactive protein. It was compared against the performance of five pulmonologists in terms of sensitivity, specificity, and others.
Key Findings:
Model Performance:
• The DES model resulted in an area under the receiver operating characteristic (ROC) curve of 0.77±0.01.
• Sensitivity: 76.2%±2.04%, 6.5% better than the average pulmonologist.
• Specificity: 63.8%±2.3%.
• Positive predictive value (PPV): 41.6%±1.2%.
• F1-score: 53.8%±1.0%.
Comparison with Pulmonologists:
• DEs model outperformed all the five pulmonologists, while sensitivity was exceptionally high, particularly for early-stage detection.
Significant Biomarkers:
• Smoking status, LDH, serum calcium levels, low sodium levels, leucocytes, neutrophil count and CRP values were the strongest predictors of LC.
The DES ML model clearly outperformed pulmonologists in detecting lung cancer with higher sensitivity and identification of critical biomarkers for risk stratification. These findings show the ability of ML in changing the current landscape of early detection strategies for improving outcomes for patients at risk of LC. ML-driven models integrated into health care could provide decision-making support to clinicians and may decrease the rate of late-stage cancer diagnosis.
Reference:
