AI-Driven Personal Risk Scoring for Insurance Underwriting

Thesis Abstract

The increasing availability of diverse personal data and advances in machine learning have positioned underwriting as a data-driven decision process, yet traditional risk models often rely on limited demographic and historical claims data, potentially overlooking nuanced individualized risk factors. This study addresses the problem of insufficient granularity and predictive accuracy in insurance underwriting by developing an AI-driven personal risk scoring framework that integrates heterogeneous data sources, including wearable device metrics, transactional behavior, health records, and social determinants, to improve risk differentiation and pricing precision while ensuring ethical and regulatory compliance. The aim is to design, implement, and validate a robust risk scoring model that produces individualized risk scores suitable for underwriting decisions, and to analyze the impact of these scores on pricing accuracy, portfolio diversification, and fairness. Specific objectives include (1) identifying a comprehensive set of predictive features across health, lifestyle, financial, and behavioral domains; (2) developing a multimodal machine learning architecture that fuses structured, unstructured, and time-series data; (3) evaluating model performance against baseline actuarial models using metrics such as AUC-ROC, Brier score, calibration plots, and net monetary value; (4) examining the model for fairness and potential biases across age, gender, and socioeconomic groups; and (5) providing policy and governance recommendations for deployment in underwriting workflows. The methodology adopts a sequential explanatory mixed-methods design. The population comprises insured individuals within a large diversified portfolio from three partner insurers. A stratified random sample of 20,000 policyholders is drawn, with a 12-month historical window for claims and a 6-month window for real-time data streams from wearable devices and digital interactions. Data collection employs (i) structured claims and policy data from enterprise data warehouses; (ii) wearable-generated metrics (e.g., activity, heart rate variability, sleep patterns) via secure APIs; (iii) health records and medication histories, subject to consent and data minimization; and (iv) socioeconomic indicators inferred from transactional and geospatial data, anonymized and aggregated to preserve privacy. Instrumentation includes a feature engineering pipeline, consent management interfaces, and model documentation artifacts to support reproducibility. Analytical methods begin with data preprocessing, feature selection using LASSO and SHAP-based importance scoring, and the development of a hybrid model architecture that combines gradient boosting for tabular features with recurrent neural networks for time-series signals and transformer-based encoders for unstructured notes. Model validation uses nested cross-validation and temporal holdout to reflect underwriting deployment. Evaluation metrics encompass discrimination (AUC-ROC, G-mean), calibration (reliability diagrams, Brier score), and decision-analytic measures (expected utility, value of information). Fairness analysis incorporates disparate impact tests, equalized odds, and counterfactual evaluation to assess subgroup consistency. Comparative analysis includes conventional actuarial models (GLMs with incremental features) and the proposed AI-driven framework. A supplementary qualitative component gathers underwriting professionals’ feedback through semi-structured interviews (n=25) to examine interpretability, governance, and rollout challenges, analyzed via thematic analysis. Expected findings indicate that the AI-driven personal risk scoring framework substantially improves predictive accuracy and calibration relative to baseline models, with AUC-ROC improvements in the range of 0.05–0.12 and reduced mispricing across high- and low-risk segments. The study anticipates that incorporating wearable and behavioral data will yield marginal gains for younger cohorts while materially enhancing discrimination for high-risk profiles when combined with health and socioeconomic indicators. Fairness analyses are expected to reveal manageable bias levels under principled consent and data minimization practices, with proposed corrective measures such as threshold adjustments and subgroup-aware calibration. Theoretical contributions include integrating information asymmetry and signal processing perspectives within insurance economics and extending fairness-aware AI governance in actuarial practice. The study contributes to knowledge by offering a validated, ethically aligned framework for AI-enabled underwriting that harmonizes accuracy with fairness and transparency, accompanied by a governance blueprint for data stewardship, consent management, and model risk mitigation. Practical implications encompass actionable guidelines for feature sourcing, model integration into underwriting workflows, risk-based pricing implications, and a roadmap for regulatory compliance and stakeholder trust. Recommendations emphasize privacy-preserving data practices, ongoing performance monitoring, explainable AI interfaces for underwriters, and staged deployment with continuous learning and robust audit trails.