The U.S. healthcare system excels at treating acute crises but fundamentally struggles with predicting chronic disease risk before it spirals into catastrophic costs. This is because current actuarial and clinical risk scoring systems evaluate patients in isolation — one row in a database, one list of diagnoses and one static score. But disease does not work that way. Healthcare risk is not isolated, it is highly contagious through shared geographic, socioeconomic, and environmental determinants. Chronic conditions such as diabetes & hypertension do not emerge from individuals in a vacuum, they propagate through communities linked by shared neighborhoods, food environments, genetics and healthcare providers. To fix broken risk prediction models, we must stop looking at patients as disconnected rows in a database and start treating them as nodes in a dynamic risk graph. My goal is to solve the problem of unpredictable chronic-disease exacerbations (such as severe asthma or heart failure) by applying iterative risk propagation across patient networks. My proposed solution models the patient population as a directed graph where each node is a patient and each edge encodes shared risk factors: overlapping diagnoses, similar demographics, geographic proximity or genetic similarities. Each patient carries a baseline risk score b derived from clinical variables such as age, BMI, and comorbidity history. Risk then propagates across this network iteratively using the following update rule, directly inspired by Schlosser’s Healthcare Risk Equilibrium Theorem (and the same eigenvector logic underlying Google’s PageRank): r(t+1) = α · A · r(t) + (1 − α) · b Here, A is the row-normalized patient adjacency matrix, and α = 0.85 controls the balance between network-propagated and baseline risk. Iteration continues until convergence. The settled solution is the dominant eigenvector of the system — a global equilibrium risk profile that reflects not just who a patient is, but who their neighbors are. I have already built a working implementation of this model in Python. The core propagation step is: new_r = alpha * np.dot(A, r) + (1 - alpha) * baseline_risk if np.linalg.norm(new_r - r) < tolerance: break To validate the prototype on synthetic data, I constructed a six-patient demonstration network with manually assigned baseline risk scores and relationship edges. Running my algorithm on this synthetic input, it converged in 20 iterations and reveals something additive models cannot: Patient 0, with a low baseline risk of 0.10, finished with a propagated score of 0.26 because their neighbors are high-risk. Patient 3, the highest baseline-risk individual at 0.60, settled at 0.37 because the network dilutes isolated risk. Even on this small synthetic example, the ranking of patients changed substantially between baseline and final scores, and that kind of shift, at scale on real data, is where preventable hospitalizations live. My approach has two concrete applications. First, it enables earlier detection of chronic disease clusters within communities. If a group of patients with shared environmental exposures begins developing metabolic disorders, the propagation algorithm raises risk scores for connected individuals before clinical symptoms appear, creating a window for preventive intervention. Second, it improves fairness in insurance pricing. Current models often penalize individuals for demographic proxies rather than true health signals. Network-based propagation identifies structural drivers of risk such as environmental gaps and treatment deserts, allowing insurers to target care rather than simply raise premiums. My long-term goal is to bring this modeling into real healthcare delivery especially in under-resourced settings where early detection is most needed. Healthcare data is not a collection of independent records. It is a network. Once we treat it that way, the mathematics of propagation and eigenvectors can help us see risk before it becomes illness.