Optimizing Donor Outreach for Blood Collection Sessions: A Scalable Decision Support Framework

André Carneiro; Pedro T. Monteiro; Rui Henriques

cs.AIMar 31, 2026

Optimizing Donor Outreach for Blood Collection Sessions: A Scalable Decision Support Framework

Authors: André Carneiro, Pedro T. Monteiro, Rui Henriques

Organizations: INESC-ID and Técnico Universidade de Lisboa Portugal

Abstract

Blood donation centers face challenges in matching supply with demand while managing donor availability. Although targeted outreach is important, it can cause donor fatigue via over-solicitation. Effective recruitment requires targeting the right donors at the right time, balancing constraints with donor convenience and eligibility. Despite extensive work on blood supply chain optimization and growing interest in algorithmic donor recruitment, the operational problem of assigning donors to sessions across a multi-site network, taking into account eligibility, capacity, blood-type demand targets, geographic convenience, and donor safety, remains unaddressed. We address this gap with an optimization framework for donor invitation scheduling incorporating donor eligibility, travel convenience, blood-type demand targets, and penalties. We evaluate two strategies: (i) a binary integer linear programming (BILP) formulation and (ii) an efficient greedy heuristic. Evaluation uses the registry from Instituto Português do Sangue e da Transplantação (IPST) for invite planning in the Lisbon operational region using 4-month windows. A prospective pipeline integrates organic attendance forecasting, quantile-based demand targets, and residual capacity estimation for forward-looking invitation plans. Results reveal its key role in closing the supply-demand gap in the Lisbon operational region. A controlled comparison shows that the greedy heuristic achieves results comparable to the BILP, with 188x less peak memory and 115x faster runtime; trade-offs include 3.9 pp lower demand fulfillment (86.1% vs. 90.0%), larger donor-session distance, higher adverse-reaction donor exposure, and greater invitation burden per non-high-frequency donor, reflecting local versus global optimization. Experiments assess how constraint-aware scheduling can close gaps by mobilizing eligible inactive/lapsing donors.

Figures & tables

Figure 1. Strategic donor outreach framework enabled by binary integer linear programming.

Figure 2. Blood donations vs. Demand in the SL region (2020). The dashed line represents 100% of demand met. Consumption of units of several blood types consistently exceed donations.

Figure 3. Distribution of last-donation recency prior to January 2020, restricted to donors who were age-eligible to donate in 2020. Donors are grouped into active, lapsing and inactive categories based on recency of their last donation; HF denotes high-frequency donors. A substantial portion of the pool had not donated for more than 10 years, indicating considerable theoretical reactivation potential.

Figure 4. Potential vs. observed demand fulfillment in the SL region (2020). The red line shows observed monthly fulfillment based on actual collections, while the green dashed line shows the simulated BILP result after targeted outreach to eligible inactive/lapsing donors. Under the assumed 5% attendance probability, the donor pool appears sufficient to close the regional supply gap.

Figure 5. Forecast-based demand targets. Panels A-B show pooled prior-year same-month carry-forward performance for CE and CPP forecasting. Panel C shows that carry-forward forecasting error is lowest in the high-volume blood groups and much higher in rare groups. Panel D shows achieved monthly coverage for trend-corrected quantile targets across blood types and quantile levels $\alpha$ : higher $\alpha$ gives a more conservative target, but the coverage gains are uneven across blood types.

Figure 6. Organic attendance forecasting for the SL operational region, evaluated on the 2020 donor-month test set after selecting an expanding training window from pre-2020 backtests and retuning the tree ensembles without class-balanced weighting. (a) ROC curves show that logistic regression, MLP, XGBoost, and LightGBM all discriminate substantially better than the historical baseline. (b) Calibration curves show that, after removing the probability distortion induced by class-balanced tree weighting, the three ML models are all reasonably calibrated, with retuned XGBoost and LightGBM slightly outperforming the MLP on Brier score. (c) Monthly aggregate organic attendance forecasts obtained by summing donor-level probabilities. The MLP remains best on monthly count accuracy (MAE 720 donors, 13.6% relative MAE), while retuned XGBoost (MAE 741, ROC-AUC 0.822, PR-AUC 0.242, Brier 0.0369) and retuned LightGBM (MAE 745, ROC-AUC 0.822, PR-AUC 0.241, Brier 0.0370) are very close.

Figure 7. Session assignment validation for the SL operational region, 2020. (a) Site-level GBR model (with top-down reconciliation): predicted versus actual annual attendance per site on log–log axes ( $r=0.983$ , MAE = 57, 387 sites); dots are coloured by annual site size. (b) Monthly organic supply: predicted returning donors (MLP aggregate, blue) and first-time donors (Holt–Winters forecast, orange) versus observed; monthly MAE = 1,033 (15.4%). The GBR determines how each monthly total is distributed across sites but is invisible at the national level shown here. (c) Cumulative distribution function (CDF) of site $\times$ month relative errors comparing three methods: GBR site model (median 31%), individual preferred-site assignment (median 72%), and proportional historical shares (median 82%).

Figure 8. Forecasting first-time donor inflow for the SL operational region. (a) Leave-one-year-out backtest MAE across 2009–2020 for four forecasting methods; legend reports overall MAE and relative MAE. Holt–Winters (MAE = 133, 17.5%) is the most accurate across the full backtest. (b) Prospective 2020 forecasts versus actual monthly counts. The COVID-19 disruption in March–April is the dominant source of forecast error across all methods. Legend reports monthly MAE and annual total for each method against the actual total of 8,406.

Figure 9. Comparison of solver approaches on the full Lisbon 2020 donor pool under a $2\times$ scaled residual demand scenario. Panels show demand fulfillment rate, number of adverse-reaction donors invited, average donor–session distance (km), average invitations per non-high-frequency donor, wall-clock runtime (s, log scale), and peak process memory (MB, log scale). The highlighted bar in each panel indicates the best-performing solver. The greedy heuristic provides highly competitive fulfillment while running substantially faster and with far lower memory usage than the exact BILP benchmark.

Appendix figures & tables6 assets

Supplementary material from the paper’s appendix.

Appendix

Figure S1. National demand fulfillment in 2020.

Figure S2. Main consumed blood components in the SL operational region data.

Figure S3. Distribution of Blood Donors by Municipality in Mainland Portugal. Choropleth map showing the number of blood donors across municipalities in mainland Portugal.

Figure S4. Annual first-time donor counts (blue, left axis) and first-time donors as a percentage of total donations (red, right axis) for the SL operational region, 2005–2020. Absolute counts declined from approximately 10,600 in 2005 and a peak of 11,500 in 2010 to 7,400–8,400 in 2016–2020, with an OLS trend of $-144$ donors per year ( $p=0.059$ ). Despite the absolute decline, the first-time share of total donations remained relatively stable at 8–10%, indicating that overall collection volumes have contracted in parallel. The uptick in 2020 (8,406 first-time donors, 10.4% share) is consistent with pandemic-driven recruitment campaigns targeting new donor populations.

Policy	Model	Mean Monthly MAE	Mean Relative MAE (%)	Mean Brier	Mean ROC-AUC	Backtest Years	Best Policy For Model
expanding	Historical Baseline	8519.2	149.2	0.0839	0.6346	4	Yes
rolling_3y	Historical Baseline	8519.2	149.2	0.0839	0.6346	4	Yes
rolling_5y	Historical Baseline	8519.2	149.2	0.0839	0.6346	4	Yes
expanding	LightGBM	37253.8	652.1	0.1532	0.8373	4	Yes
rolling_3y	LightGBM	38451.5	672.6	0.1609	0.8382	4	No
rolling_5y	LightGBM	38567.0	675.0	0.1609	0.8382	4	No

Appendix

Table S1. Pre-2020 training-window backtest summary for the organic attendance forecasting workflow. Mean metrics are averaged over evaluation years 2017–2020. The policy-sweep stage compared the historical baseline, logistic regression, XGBoost, and LightGBM; the MLP was introduced only in the subsequent tuning stage after the window policy had been fixed.

Method	Site-level $r$	Site-level MAE	Monthly MAE	Median rel. err (%)
Proportional shares	0.971	80	1,033	82
Individual pref. site	0.978	89	1,197	72
GBR site model (reconciled)	0.983	57	1,033	31

Appendix

Table S2. Comparison of session-assignment methods for the SL operational region, 2020. Three allocation strategies are evaluated on 387 active sites: (i) proportional historical shares, (ii) individual preferred-site assignment (MLP + ghost-site remapping), and (iii) the site-level GBR model. The GBR model achieves the lowest site-level MAE and median relative error, making it the preferred method for capacity estimation. The individual assignment, while less accurate at the site level, provides per-donor site linkages required by the invitation optimizer.

Healthcare

Healthcare AI develops and evaluates models and systems for clinical care, medical research, public health, and healthcare operations.

See all 616 papers

A General Framework for Budgeted Threshold Incentives on Request
Sep 24, 2026Zhuolin Wu, Chengrui Zhu, Wenhua Nie +9Causal Effect Estimation Online Resource Allocation
Meituan · National Taiwan University · Tsinghua University +1
Generative Frontier Planning for Adaptive Peer-Referral Recruitment under Covariate-Dependent Arrivals
Jun 6, 2026Lingkai Kong, Hezi Jiang, Andrew Ma +3Adaptive Sampling Model-Based Planning
Harvard University
Online Allocation with Unknown Shared Supply
May 8, 2026Tzeh Yuan Neoh, Davin Choo, Mengchu Yue +1Online Resource Allocation Sequential Decision Making
Harvard University

Sep 24, 2026J/K move · Enter open · S save

A General Framework for Budgeted Threshold Incentives on Request

Zhuolin Wu, Chengrui Zhu, Wenhua Nie, Kenny Ye Liang+8

Meituan · National Taiwan University · Tsinghua University+1

Causal Effect Estimation Online Resource Allocation

Open paper AskPDF arXiv

On-demand delivery platforms pay riders through incentive activities whose tiers are set from recent completions of riders with a similar history. Operators request such plans for changing periods, rider populations, payment rules and budgets, often for holidays or bad weather, where randomized trials are scarce and take months to collect. We present a request-driven framework that composes four stages (conditional prediction, population reduction, trajectory integration and budget allocation) through seven replaceable modules that exchange conditional trajectory laws, whose award probabilities and award-marked moments give payment and uplift for any activity rule. A response-correction step reweights trajectories from abundant no-offer history to match the moments of a short pilot. We prove that, on a fixed plan menu and given the stage errors, the end-to-end value loss is bounded by the sum of four stage terms, and that for every stage there are instances on which omitting it leaves an error floor the others cannot remove. On 3,000 riders over 45 weekly origins, all 127 windows of a week are answered 11.04x faster with identical scenarios and at most 0.92% value lost by the allocation. On 24 new controlled response laws, the response correction with a one-week pilot lowers regret by 51.2% relative to a trial with the same nominal randomized rider-weeks, and a four-week pilot with exact summation comes within +0.007 of an 18-week trial. In registered studies where windows, populations, rules and binding budgets change from request to request, the framework's regret is below that of a trial with the same nominal rider-weeks and below dose interpolation of the same pilot data, and reusing its one-off preparation answers 60 requests 14.1x and 2.70x faster with identical answers. Against a nine-offer trial fitted with the framework's own dose curve, one-week regret is 0.055 lower.

Abstract

Healthcare

Explore similar work

A General Framework for Budgeted Threshold Incentives on Request

Generative Frontier Planning for Adaptive Peer-Referral Recruitment under Covariate-Dependent Arrivals

Online Allocation with Unknown Shared Supply