MATHx Lab

Research

MATHx advances human health by developing and applying rigorous computational models that transform complex biomedical data into actionable insights, enabling earlier diagnosis, precise risk stratification, and scalable clinical impact across populations.

Areas of Focus

Comparative Effectiveness and Outcomes

The Kang Lab uses computer-based simulation modeling to understand how medical imaging is best used to improve patient health outcomes, synthesizing the complex tradeoffs of benefits and harms for different tests and treatments.

  • We are constructing mathematical models for several types of neoplasia to assess risk-based management, including kidney and bladder tumors, oral squamous cell carcinoma, and prostate cancer.
  • We have guided use of decision science methods to assess and calibrate deep learning diagnostic models for oral lesions and breast cancer screening.
  • Our group has demonstrated how competing risk analysis improves treatment selection for incidental kidney tumors; potential kidney functional decline after surgery must be weighed against oncologic risks for effective and personalized treatment selection.
  • We are assessing translation of modeling results to patient care through a web-based decision support tool for patients who are deciding on treatment for small kidney tumors.
  • We perform cost effectiveness analyses. This form of economic evaluation incorporates life expectancy or quality-adjusted life expectancy as well as cumulative costs of the strategies being compared. For example, I have assessed personalized diagnostic pathways using prostate MRI in men diagnosed with low-risk prostate cancer considering active surveillance. Thresholds for repeat biopsy in men with preferences to limit the number of unnecessary lifetime biopsies may be set to a higher MRI score (with more likely clinically significant tumor), instead of the current threshold of an “equivocal” result. Annual MRI with this higher threshold for repeat biopsy both limits the number of unnecessary biopsies and is cost-effective. This work has been featured in urology and oncology media reports, strengthening clinical relevance and further collaboration.
Figure illustrating treatment pathways for small renal tumors represented in the model.

Treatment pathways for small renal tumors represented in the model. ∗Renal function risk assessment guides the selection of a nephron-sparing therapy (partial nephrectomy or percutaneous ablation) by using tumor nephrometry score and chronic kidney disease stage (21). Partial nephrectomy is selected for patients with normal renal function, and for stage 2 or 3a chronic kidney disease and nephrometry score of 6 or less; otherwise, percutaneous ablation is used. aAny renal cell carcinoma (RCC) subtype leads to initial treatment with nephron-sparing treatment in the base case. bActive surveillance includes CT scan every 6 months for 1 year followed by yearly CT scan to assess for growth; increase in size of more than 3 mm in 1 year triggers nephron-sparing treatment according to tumor anatomy and patient comorbidities. c Prediction of papillary subtype of RCC by applying diagnostic accuracy of contrast enhanced MRI. pRCC = papillary RCC. ( Kang SK, Huang WC, Elkin EB, Pandharipande PV, Braithwaite RS. Personalized treatment for small renal tumors: decision analysis of competing causes of mortality. Radiology. 2019 Mar;290(3):732-43. https://doi.org/10.1148/radiol.2018181114)

Evidence Synthesis for Diagnostic Test Performance

We evaluate the evidence basis on diagnostic tests using systematic review and meta-analysis.

  • Through hierarchical modeling, we have examined which specific technical factors of current MRI (e.g., contrast agent, with or without particular sequences) yield better diagnostic accuracy, and whether different threshold values for test positivity account for the apparent heterogeneity in the sensitivities and specificities reported across the literature.
  • We lead meta-analysis of diagnostic performance studies of prostate MRI, renal mass characterization, and MRI screening for hepatocellular carcinoma. We have also applied meta-analysis to examine discrepancy rates and clinical impact of secondary interpretations of imaging. Several of these meta-analyses have been cited in broadly applied clinical practice guidelines in the U.S. and internationally.
Figure illustrating a summary ROC (SROC) curve of the accuracy of MR imaging detection of small hepatocellular carcinoma (HCC).

Summary ROC (SROC) curve of the accuracy of MR imaging detection of small hepatocellular carcinoma (HCC). The confidence region (conf.region) represents the ellipsoid 95% confidence region in summary ROC space for the summary point estimate of diagnostic performance. (Kierans AS, Kang SK, Rosenkrantz AB. The diagnostic performance of dynamic contrast-enhanced MR imaging for detection of small hepatocellular carcinoma measuring up to 2 cm: a meta-analysis. Radiology. 2016 Jan;278(1):82-94. https://doi.org/10.1148/radiol.2015150177)

Improving Healthcare Delivery for Incidental Findings

Our team uses a mixed-methods approach to study disruptions in follow-up care for several types of common radiological incidental findings, with the goal of developing toolkits to improve delivery of guideline-concordant care for incidental lesions.

  • We have developed an electronic medical record tracking tool of incidental lung nodules with lesion risk category and follow-up recommendations with due dates. We have conducted a large-scale analysis of incidental lung nodules, comparing follow up rates before versus after implementation of a standardized reporting system that provides the specific recommendation in the report and tracks lesions. Over 10,000 patients’ incidental lung nodules have been tracked for appropriate follow up. We have also conducted a prospective pilot study on patient notification.
  • We are developing a surveillance system that accounts for variables impacting disparities in outcomes of high-risk lung nodules, for example in census tracts with higher social vulnerability index. We have used a large language model (GPT-4) to facilitate rapid summarization of follow up needs.
Figure illustrationg the percentage of nodules with appropriate timing of follow-up in preintervention cohort and postintervention cohort when follow-up occurred,

Percentage of nodules with appropriate timing of follow-up in the preintervention (2014) cohort and postintervention (2018) cohort when follow-up occurred, showing a significantly higher follow-up for all incidental lung nodules (P < .001) but a nonsignificant difference in nodules ≥ 9 mm (P = .178) (Bagga B, Fansiwala K, Thomas S, Chung R, Moore WH, Babb JS, Horwitz LI, Blecker S, Kang SK. Outcomes of incidental lung nodules with structured recommendations and electronic tracking. Journal of the American College of Radiology. 2022 Mar 1;19(3):407-14. https://doi.org/10.1016/j.jacr.2021.09.046)

Back to top