New AI Tool Offers Hope for Early Detection of Cancer Cachexia
Cancer cachexia is a serious condition that impacts up to 80% of people with advanced cancer. It leads to extreme weight loss, muscle wasting and fatigue. Unfortunately, simply improving nutrition isn't enough to reverse cachexia.
Despite being common, cachexia is hard to diagnose. It can significantly affect quality of life, treatment outcomes and survival rates, and often it isn't recognized until it has reached an advanced stage.
The complexity of cachexia adds to the challenge. Its symptoms can be similar to those caused by cancer treatments or other issues like malnutrition. Additionally, inconsistent weight tracking, cultural sensitivities around discussing weight loss and a lack of standardized diagnostic tools make early detection even more difficult.
Now, Moffitt Cancer Center researchers have developed an artificial intelligence (AI)-driven biomarker model that could change how cachexia is identified. The tool integrates imaging scans with routine clinical data to provide a more accurate and timely diagnosis. Findings from a study validating the tool were presented at the American Association for Cancer Research annual meeting.
Sabeen Ahmed
“Cancer cachexia is a devastating complication that often goes undetected until it’s too late,” said Sabeen Ahmed, a graduate student involved in the study. “Our AI-driven approach combines multiple sources of clinical data to identify cachexia earlier, enabling interventions that can slow muscle wasting and improve metabolic function.”
The AI model analyzes CT scans to quantify skeletal muscle mass using an algorithm trained on annotated images. This data is then combined with laboratory results, electronic medical records and other clinical information to predict whether a patient has or will develop cancer cachexia. By combining imaging features derived from the CT data with clinical data, which includes patient demographics, height, weight, body mass index and cancer stage, the model accurately identified cachexia for 77% of pancreatic cancer patients enrolled in the study. The addition of laboratory findings (albumin, neutrophils, lymphocytes, creatinine and blood urea nitrogen) improved this accuracy to 81% and further to 85% with the incorporation of structured clinical notes.
Survival analysis using multimodal data for patients with pancreatic, colorectal and ovarian cancers demonstrated significant improvements in accuracy of 6.7%, 3.0%, and 1.5% respectively, compared to using the clinical data alone.
Ahmed emphasized that integrating diverse data sources is key to the model’s success. “Each data type offers a puzzle piece,” she explained. “By combining them, our model uncovers hidden patterns that single tests might miss, providing a more accurate diagnosis.”
The researchers also highlighted the reliability of their approach. The model’s skeletal muscle measurements differed by just 2.48% from manual assessments conducted by expert radiologists. Additionally, the approach provides an estimate of the model’s confidence in its measurements, helping to flag model measurements that are likely to deviate significantly from manual assessments. These features demonstrate its potential for real-world use.
While the study focused on a limited number of cancer types and relied solely on CT scans for imaging data, researchers are optimistic about expanding its applications. Incorporating PET or MRI scans could enhance accuracy, while adapting the model for other cancers may broaden its impact.
“Our findings highlight how machine learning can transform cancer care,” Ahmed said. “By integrating diverse clinical data sources, we can generate insights that empower personalized treatment plans and improve patient outcomes.”
