Eileen Rakovitch

University of Toronto
Country flagCanada

Research Interests

ApplyKite TwinLearns as you apply

Make your next move.

Prepare your application

Review drafts before you send them.

Explore related searches

Articles (6)

Year:

2026

Background Although DCIS has a relatively low recurrence rate, many patients still receive adjuvant radiotherapy or endocrine therapy, raising concerns about overtreatment. Reliable biomarkers are therefore needed to predict an individual patient’s risk and guide treatment decisions. Recent studies suggest that the composition of the tumour-associated stroma (TAS) affects progression and outcome, highlighting TAS-derived biomarkers as promising candidates for further investigation. Methods We trained AI models for cell and tumour segmentation using whole slide digital pathology images acquired as part of a retrospective cohort study. We investigated the effects of cell density within both the tumour and the TAS to determine how they correlated with recurrence in the ipsilateral breast. Results We found that the concentration of DCIS lesions on the slide and the density of mitotic figures inside the TAS region were significantly associated with recurrence risk. Additionally, we found some predictive value in the lymphocyte and red blood cell densities in different tumour regions. Stromal composition was shown to associate with recurrence risk, and density-based biomarkers were identified and used to cluster patients into phenotypes with significantly different risk profiles. Conclusion Our findings highlight the prognostic relevance of stromal composition in DCIS, and we identify novel density-based biomarkers that can be used to identify patients who are more likely to experience a local recurrence after breast-conserving surgery alone. These results may aid in developing future risk-stratification tools for breast cancer patients, thereby reducing overtreatment and improving patient care.

Year:

2026

Formalin-fixed paraffin embedded (FFPE) samples are challenging to profile using existing high-throughput sequencing technologies, including RNA-seq. This difficulty primarily arises from the degradation of nucleic acids, a problem that becomes particularly acute with samples stored for extended periods. FFPE-derived RNA-seq (fRNA-seq) data have a high rate of transcript dropout, a property shared with single cell RNA-seq. Transcript counts also have high variance and are prone to extreme values. We introduce the PaRaffin Embedded Formalin-FixEd Cleaning Tool (PREFFECT), a probabilistic framework for the analysis of fRNA-seq data. PREFFECT uses generative models to fit distributions to observed expression counts while adjusting for technical and biological variables. The framework can exploit multiple expression profiles generated from matched tissues for a single sample (e.g., a tumor and morphologically normal tissue) in order to stabilize profiles and impute missing counts. PREFFECT can also leverage sample-sample adjacency networks that assist graph attention mechanisms to identify the most informative correlations in the data. We demonstrate how PREFFECT uses this information to more accurately compute downstream analyses such as sample clustering in several datasets.

Year:

2025

PURPOSE Ductal carcinoma in situ (DCIS) is routinely treated with adjuvant radiotherapy (RT) after breast-conserving surgery (BCS). The inability to accurately estimate an individual's risk of local recurrence (LR) and invasive LR using clinicopathologic factors (CPF) contributes to the overtreatment of DCIS. We examined the impact of the 12-gene DCIS Score (DS) and the 21-gene Recurrence Score (RS) on the accuracy of predicting LR and invasive LR. METHODS A population-based cohort diagnosed with pure DCIS treated with BCS ± RT from 1994 to 2003 was used. All patients had expert pathology review and assessment of the DS and RS. Predictive models (CPF alone, DS + CPF, and RS + CPF) were developed using multivariable Cox regression analyses to predict 10-year LR and invasive LR risks. Models were evaluated on the basis of c-statistic, –2log likelihood estimate (–2LLE), and Akaike information criterion. Calibration was performed using bootstrap resamples, with replacement. RESULTS The cohort includes 1,226 women treated with BCS; 712 received RT. 194 women (15.8%) experienced ipsilateral LR as a first event; 112 were invasive. Models including the DS or RS performed better in predicting the 10-year risk of LR compared with models on the basis of CPF alone with excellent calibration. The two molecular-based models also performed better in predicting invasive LR compared with the CPF model but the model incorporating the RS did not perform better in the prediction of invasive LR compared with the DS-based model. CONCLUSION Models incorporating the DS or RS more accurately predicted the 10-year risk of LR and invasive LR after BCS compared with models on the basis of CPF alone. Inclusion of the RS, compared with DS, did not improve the prediction of the 10-year risk of invasive LR.

Year:

2024

Background: A twelve-gene molecular expression assay (DCIS score) may help guide radiation oncology treatment under specific circumstances. We undertook a study to examine radiation oncologist (RO), surgeon, and decision maker views on implementing the DCIS score in practice for women with low-risk DCIS. Methods: We conducted a qualitative study involving telephone interviews that were audio-recorded and transcribed. Two researchers conducted a thematic analysis of transcripts. Results: Twenty-eight individuals (ROs, breast cancer surgeons, and cancer policy decision makers) were invited to participate; 22 out of the 28 people (79%) agreed. The final sample included 20 participants: 11 of 13 (85%) ROs, 5 of 7 (71%) surgeons, and 4 of 8 (50%) decision makers. Most ROs expressed concerns about overtreatment but could not predict with certainty which low-risk patients could safely avoid radiation. The DCIS score was viewed as contributing valuable personalized risk information as part of treatment decision making that included clinicopathological factors and women’s preferences. Future implementation would require guidelines with input from the oncology team. Conclusions: ROs had concerns about the overtreatment of women with DCIS, but lacked the tools to reliably predict which women could safely avoid radiation. By providing oncologists and women with personalized tumor information, the DCIS score was an important component of treatment decision making.

Year:

2024

Ductal carcinoma in situ (DCIS), especially in the era of mammographic screening, is a commonly diagnosed breast tumor. Despite the low breast cancer mortality risk, management with breast conserving surgery (BCS) and radiotherapy (RT) is the prevailing treatment approach in order to reduce the risk of local recurrence (LR), including invasive LR, which carries a subsequent risk of breast cancer mortality. However, reliable and accurate individual risk prediction remains elusive and RT continues to be standardly recommended for most women with DCIS. Three molecular biomarkers have been studied to better estimate LR risk after BCS—Oncotype DX DCIS score, DCISionRT Decision Score and its associated Residual Risk subtypes, and Oncotype 21-gene Recurrence Score. All these molecular biomarkers represent important efforts towards improving predicted risk of LR after BCS. To prove clinical utility, these biomarkers require careful predictive modeling with calibration and external validation, and evidence of benefit to patients; on this front, further research is needed. Most trials do not incorporate molecular biomarkers in evaluating de-escalation of therapy for DCIS; however, one—the Prospective Evaluation of Breast-Conserving Surgery Alone in Low-Risk DCIS (ELISA) trial—incorporates the Oncotype DX DCIS score in defining a low-risk population and is an important next step in this line of research.

Year:

2023

Collaborators (4)

Michael T. Hallett

Professor

Western University

CANADA
View details

Katarzyna Joanna Jerzak

University of Toronto

CANADA
View details

Anne Martel

Professor

University of Toronto

CANADA
View details

Vanessa Dumeaux

Assistant Professor

Western University

CANADA
View details