Computational prediction of multidisciplinary team decision-making for adjuvant breast cancer drug therapies: a machine learning approach

The central findings of this study are two-fold. First, a machine learning-based approach is useful for predicting MDT decisions about adjuvant drug therapies in early breast cancer patients; to the best of our knowledge, this is the first systematic analysis of predictive modelling of the MDT outcome in breast cancer. Second, unlike adjuvant hormone or trastuzumab MDT decisions, adjuvant chemotherapy MDT decisions differed significantly from guideline-based decisions, suggesting that additional non-clinicopathologic variables impact upon expert advice in the adjuvant chemotherapy context. These findings could reflect chemotherapy-specific decision variations due to divergences in patient preference, cultural or socioeconomic differences, and resource availability. Since machine learning remained predictive of MDT decisions, we speculate that future work may succeed in identifying these important missing data, and thus help to understand this discrepancy better.

For early breast cancer patients, oncologists and their professional colleagues must determine the most appropriate adjuvant therapy. A multidisciplinary approach is important in making decisions about adjuvant treatments after a surgical resection with curative intent; the goals of recurrence reduction (deferral, cure) must be carefully weighed against the toxicity, cost, inconvenience and other detriments to patient quality of life. Although MDT opinions on whether a patient should undergo toxic treatment can be contentious between experienced clinicians, the benefits of a multidisciplinary approach clearly reduce breast cancer-specific mortality [18].

The goal of our modelling differs from prognosis-based decision aids such as Adjuvant! and the PREDICT Tool [19, 20], where the primary goal of these tools is to estimate benefits for a given level of risk for recurrence and/or death. A practical objective of our study is therefore to assess the feasibility of predicting the actual MDT outcome, which captures the practical aspects other than solely the survival considerations of a patient.

We found that the machine learning models were high discriminative of the outcome variables, with the predictive accuracy consistently achieved at a clinically useful level. The internal validity was demonstrated by thorough cross-validation evaluations. Further studies at an external centre would clarify its clinical utility. We expect our analytic approach could also predict MDT recommendations for other treatment modalities such as surgery and radiotherapy, as well as assist in decision-making for patients suffering metastatic disease.

Our analysis is strengthened by a comprehensive survey of classifiers with distinct inference techniques; the comparative design has allowed determination of the best algorithm for each task. The alternating decision tree algorithm outperformed other classifiers for predicting MDT decisions about endocrine and trastuzumab therapies; on the other hand, the bootstrap-aggregated ripple-down-rules classifier was superior for predicting adjuvant chemotherapy decisions. We conclude from this that a tree-based approach resembles more closely how experts make actual decisions in a collaborative environment. Conversely, both generative and discriminative probabilistic methods (such as naïve Bayesian classifier and multivariate logistic regression) did not perform as well as tree-based classifiers; one explanation for this may be that these algorithms were compromised by strong co-linearity between certain variables. Aggressive feature selection may thus be required to optimise their performance.

For decisions about adjuvant chemotherapy, significant discrepancies were apparent between MDT decisions and the two international guidelines. Guideline-driven individualisation of treatments may thus prove challenging; factors such as treatment toxicity, performance status, quality of life, psychological well-being, and patient’s perception of treatment efficacy can strongly influence the treatment decision [8, 21–23], but such nuances are poorly captured by practice guidelines. Consequently, while evidence-based guidelines are designed to suit the majority of patients, our study highlighted the importance of individualised, patient-centred assessments as per best MDT practice. Identification of putative underlying non-clinicopathologic variables through a machine learning approach could help to elucidate how clinicians arrive at MDT decisions about adjuvant chemotherapy for early breast cancer.

A potential use of our modelling approach is to allow estimation of decision consistency within a cancer MDT. Intuitively, the most accurate model also indicates how well an MDT outcome can be predicted using the same clinical and pathological characteristics. A comparative evaluation of multiple models hence provides an objective mean for which the auditing of decision quality can be conducted within and/or between cancer centres.

Several applications are made possible by the machine learning approach described here. First, the most predictive classifier(s) can be packaged into a site-specific decision support system to help real-time decision making in a MDT, which has the potential to enhance the decision making process by considering local resource constraints compared with using an external guideline. The use of a computerised decision support can also improve uptake of evidence-based care [24]. Second, a reliable model should enable transfer of knowledge to smaller or less experienced centres, for example, in remote or rural settings, thus permitting early triage or referral of complex cases. Third, the decision about individual cases can be compared across different centres, which would otherwise not be feasible to do.

It is important to acknowledge that our study has several limitations. First, our data did not fully record the sequencing of treatment modalities, investigations, or chemotherapy regimens, which would otherwise allow us to fine-tune the predicted recommendations. Second, final decisions after patient review by medical oncologists (i.e., as distinct from the “intention to treat” recommendations recorded in MDTs) were not always available to us; we expect that these final treatment outcomes are modified by additional elements of patient preference. Third, survival benefits were unable to be quantified from our non-randomised (retrospective) data, since early breast cancer patients have a relatively good prognosis; a very large sample size with lengthy follow up would be required to draw meaningful conclusions on survival benefit. Fourth, our data did not fully record all administrative confounders, such as absence of a specific expert(s) from the MDT, delays in assessment, or attendance of the meeting. It is known that the team, social, and information factors do influence decisions made in a MDT [25]. A prospective study aiming to address these issues would thus be important to support solid models in the future. Finally, the present study represent only the expertise from a single cancer centre and hence may not reflect clinical practice elsewhere, though supervised learning approach can be readily extended to aggregate expertise from multiple centres. Despite the limitations, the demonstrated predictive accuracy of our study supports the future research studies of the machine learning model in a clinical setting.