Cambridge Personal Statement Example: Medicine to health data science (Score 93)

Unplanned thirty-day hospital readmissions represent one of the most measurable and policy-tractable inefficiencies in health systems managing ageing populations. In China, where the proportion of adults aged sixty-five and over is projected to exceed twenty percent by 2035, multimorbidity — the co-occurrence of two or more chronic conditions — substantially elevates readmission risk, yet existing prediction tools have been developed predominantly on Western administrative datasets and validated in single-payer institutional contexts that differ structurally from China's tiered hospital system. This proposal asks: among older adults with multimorbidity discharged from secondary and tertiary hospitals in urban China, which combinations of clinical, socioeconomic, and care-pathway factors most reliably predict thirty-day readmission, and what do those predictors imply for discharge planning policy? The question is bounded by three design choices made at the outset. First, the outcome is restricted to unplanned readmissions within thirty days, a window that is both clinically meaningful and administratively traceable. Second, the population is limited to patients aged sixty and above with at least two recorded chronic diagnoses at index admission, targeting the subgroup where readmission rates are highest and where policy intervention is most warranted. Third, the analytical frame is explicitly policy-oriented: the goal is not to build a clinical risk score for bedside use but to identify system-level and discharge-pathway variables that health authorities can act on, a distinction that shapes both the variable selection strategy and the interpretation of findings. The readmission prediction literature has grown substantially since the introduction of the Hospital Readmissions Reduction Program in the United States, and several systematic reviews have catalogued the predictive performance of models built on claims data, electronic health records, and composite indices such as the LACE score. Three gaps remain relevant to this proposal. Most validated models treat comorbidity burden as a single composite score rather than examining which specific disease dyads or triads drive readmission risk, an aggregation that obscures actionable clinical patterns. Socioeconomic and care-continuity variables — including insurance type, distance to primary care, and whether a structured discharge plan was documented — are inconsistently included, despite evidence from lower-middle-income settings that these factors carry independent predictive weight. Most directly pertinent here, no published model has been trained and validated on linked inpatient and outpatient administrative data from China's New Rural Cooperative Medical Scheme or Urban Employee Basic Medical Insurance populations, meaning that existing tools cannot be assumed to transfer without recalibration. The proposed study addresses this third gap directly and contributes to the first two by using a variable selection approach that preserves disease-pair interactions and explicitly tests care-continuity proxies. The study will use a retrospective cohort design drawing on linked hospital discharge records, outpatient encounter data, and insurance claims from a provincial health information platform in China covering 2019 to 2023. The target sample is all index admissions of patients aged sixty and above with two or more chronic condition codes recorded at discharge, excluding planned readmissions identified through elective procedure codes and patients who died during the index stay. The primary outcome is unplanned readmission within thirty days of discharge. Variable construction will follow a pre-specified coding protocol. Clinical variables will include primary and secondary diagnosis codes mapped to ICD-10 chapters, Charlson Comorbidity Index components disaggregated to individual conditions, and length of stay at index admission. Care-pathway variables will include whether a follow-up outpatient appointment was registered within seven days of discharge, the number of distinct prescribing providers in the ninety days before admission, and admission route. Socioeconomic proxies will be derived from insurance scheme type and residential district-level deprivation quintile, the finest geographic resolution available in the administrative data. The primary analytical approach will be penalised logistic regression with elastic net regularisation, chosen because it produces interpretable coefficient estimates suitable for policy communication while handling the moderate multicollinearity expected among comorbidity indicators. A gradient-boosted tree model will be estimated in parallel as a performance benchmark. Model discrimination will be assessed by the area under the receiver operating characteristic curve with bootstrapped confidence intervals, and calibration will be assessed using the Hosmer-Lemeshow statistic and calibration plots across deciles of predicted risk. Internal validation will use five-fold cross-validation. If the provincial dataset permits a temporal split — pre-2022 training, 2022 to 2023 validation — that will be the preferred validation strategy, as it better reflects prospective deployment conditions. Access to the provincial administrative dataset is being pursued through a formal data-sharing agreement between the applicant's home institution and the provincial health bureau; a letter of intent has been prepared and the data governance process is expected to conclude before the MPhil begins. The dataset is de-identified at source under Chinese personal information protection regulations, and no re-identification of individuals is required by the analysis plan. An application to the relevant departmental ethics panel will be submitted in the first term, covering secondary data use and the data transfer protocol. The proposed timeline spans three terms. The first term will be devoted to data access finalisation, variable coding protocol pre-registration, and a targeted literature review focusing on readmission prediction in East Asian health systems. The second term will cover data cleaning, cohort construction, and primary model estimation. The third term will be allocated to validation analyses, sensitivity analyses testing alternative readmission windows of seven and ninety days, and write-up. The scope is deliberately constrained to a single province and a four-year window to remain achievable within the MPhil timeframe; the study does not attempt national representativeness and the write-up will state this boundary explicitly. The main feasibility risk is delayed data access; a contingency plan exists in the form of a publicly available Chinese hospital discharge dataset that would allow the core modelling approach to be demonstrated on a smaller sample while provincial access is confirmed. The MPhil in Health Policy at Cambridge provides the methodological and policy-analytic environment this project requires. The programme's emphasis on quantitative health systems analysis and its connections to units working on health services research align directly with the study's orientation toward system-level inference. The applicant's background in clinical audit and epidemiology, including prior work applying administrative data methods and a placement producing a policy briefing on discharge pathway design, provides the technical preparation needed to execute the proposed analysis. The applicant has identified potential supervisors whose published work on hospital performance indicators and multimorbidity in ageing populations is directly relevant, and preliminary contact will be made upon application. This study will produce a readmission prediction model calibrated to a Chinese administrative data environment, a variable importance analysis identifying which care-pathway and socioeconomic factors carry independent predictive weight after clinical adjustment, and a policy-oriented interpretation of findings framed around discharge planning decisions that health bureaux can act on. The contribution is modest in scope and honest about its geographic and temporal limits, but it addresses a specific gap in the transferability of readmission models to non-Western health system contexts and offers a replicable analytical template for other provincial datasets.