A mathematical and computational framework to predict the time to and recovery from osteoporosis via serial DEXA scans: a proof-of-concept model for digital decision support Authors: P. Scully, Azrin Muslim, T. Sheehy, A. Costelloe, N. Asri, D. Lyons Published: 2026 ABSTRACT BackgroundDual-energy X-ray absorptiometry (DEXA) is the diagnostic standard for osteoporosis, yet its serial data remains underutilised in predictive analytics. To our knowledge, no published model provides explicit age-based predictions of osteoporosis onset or recovery using serial DEXA T-score trajectories. This proof-of-concept study describes a deterministic mathematical framework for predicting time to osteoporosis (TTO) and time to exit osteoporosis (TEO), defined as the age at which a patient’s T-score trajectory reaches or exits the threshold of − 2.5.MethodsWe developed two deterministic algorithms converting serial hip DEXA T-scores into age-based predictions: a two-point slope algorithm (TTOc) and a multipoint least-squares regression (TTOt). The algorithms were evaluated on 200 patients drawn from an institutional DEXA database using a pre-specified stratified random-sampling rule (50 patients each with 2, 3, 4, and ≥ 5 scans; seed 42). Three validation analyses were performed: (i) onset-age prediction against observed age of first osteoporotic reading using mean absolute error (MAE) and Bland-Altman analysis; (ii) prospective T-score prediction using scans 1 to N − 1 to predict scan N; and (iii) stability analysis examining how prediction intervals narrowed with increasing scan count. 95% prediction intervals were computed for TTOt by inverse prediction.ResultsOf 200 patients, 19 experienced observable crossing into osteoporosis and 4 experienced recovery during follow-up. For the 8 patients with sufficient pre-crossing data, TTOc predicted observed onset age with MAE 5.67 years (mean bias + 1.42; 95% limits of agreement − 16.0 to + 18.8). TTOt gave MAE 7.33 years (bias + 4.73; LoA − 14.2 to + 23.7). For prospective T-score prediction across 300 prediction points in 150 patients, TTOt produced lower MAE than TTOc on prospective T-score prediction (0.385 vs. 0.533 T-score units). Stability analysis demonstrated marked narrowing of TTOt prediction intervals with increasing scan count (mean PI width: 654 years at 3 scans, 202 years at 4 scans, 69.5 years at 5 scans). The two algorithms showed complementary behaviour: TTOc produced lower onset-age error in the validation subset (n = 8), while TTOt produced lower error for next-scan T-score extrapolation.DiscussionThe framework translates static DEXA outputs into patient-specific age-based predictions using mathematics that any reader can inspect on one page. TTOc and TTOt showed distinctly complementary behaviour. The two-point method produced lower onset-age error, while regression produced lower error for next-scan T-score prediction. The stability analysis further establishes a minimum data requirement for reliable interval estimation, with prediction intervals narrowing sharply between three and five scans. External validation, confounder adjustment, and evaluation against clinical endpoints are the necessary next steps before the framework could inform deployed decision support tools.ConclusionsSerial DEXA T-scores can be converted into interpretable, age-based threshold predictions using simple deterministic algorithms, with regression-based prediction error within an order of magnitude of the intrinsic measurement noise of DEXA itself. The framework is transparent by construction, providing a concrete methodological foundation on which future clinical decision support tools for bone health can be built and benchmarked.Supplementary InformationThe online version contains supplementary material available at 10.1186/s12911-026-03565-3. Introduction Osteoporosis is the most common metabolic bone disease and is characterised by impaired bone strength and increased fragility fracture risk [1]. In the United States, approximately 10 million adults aged 50 years or over have osteoporosis, and an additional 43 million are at risk due to low bone mass [2]. With global ageing populations, the burden of osteoporosis and osteoporotic fractures is expected to rise further. Bone mineral density (BMD) testing is central to diagnosis, risk prediction, and monitoring. Among the available methods, dual-energy X-ray absorptiometry (DEXA) of the spine and hip is the accepted diagnostic standard and the most reliable technique for longitudinal assessment [3]. Osteoporosis is defined by the World Health Organization (WHO) as a T-score of − 2.5 or lower at the hip or spine [4]. The International Society for Clinical Densitometry (ISCD) recommends use of the Third National Health and Nutrition Examination Survey (NHANES III) young-adult Caucasian reference database for hip T-scores across ethnic groups, with sex-specific reference data [5]. While DEXA remains the diagnostic standard for quantifying BMD, its potential for longitudinal prediction is underexplored. Conventional reporting provides a numerical T-score, a statistical measure of deviation from peak bone mass, but offers no intuitive sense of temporal progression; when a patient is likely to become osteoporotic, or conversely, when improvement is expected under therapy. Existing tools such as FRAX estimate fracture probability but do not forecast the time to threshold. This temporal gap leaves clinicians without a way to visualise or communicate disease trajectories in age-based terms. The present proof-of-concept study introduces a simple algorithmic framework that transforms serial DEXA data into interpretable age-based predictions: time to osteoporosis (TTO) and time to exit osteoporosis (TEO). The aim is to establish the feasibility of reframing densitometry as a time-based metric and to examine the behaviour, boundaries, and limitations of the simplest deterministic formulations of this idea. To our knowledge, no published model directly converts serial DEXA data into explicit time-to-threshold predictions using deterministic slope and intercept formulations. This work therefore offers an initial framework for such trajectory-based estimation, presented as a proof-of-concept with a focus on interpretability, quantitative validation against observed outcomes, and explicit uncertainty quantification. Methods Study design and data source We performed a retrospective proof-of-concept study using anonymised DEXA data from the Clinical Age Assessment Unit, University Hospital Limerick. The cohort comprised 200 adult patients who had undergone two or more hip DEXA scans between November 1999 and November 2024. Inclusion criteria and cohort selection rule Eligibility criteria were pre-specified: (a) age ≥ 18 years at first scan; (b) two or more hip DEXA scans on record; and (c) complete total-hip T-score data at all included scans. Two data quality filters were applied at the patient level: scan dates had to fall within 1990 to 2025, and a patient’s recorded ages across scans had to be non-decreasing (with a 0.5-year tolerance for same-year rounding); patients with internally inconsistent age records were excluded. Eligible patients were stratified by total number of hip DEXA scans available into four strata: exactly 2 scans (pool: 11,945 patients), exactly 3 scans (pool: 5,675), exactly 4 scans (pool: 2,746), and five or more scans (pool: 3,025). Within each stratum, 50 patients were drawn at random using a fixed seed (seed = 42) for reproducibility. For patients in the five-or-more-scan stratum, only the first five scans per patient were used for analysis. The selection rule was fixed before any TTO/TEO computations were performed. All scans were performed on a Lunar Prodigy densitometer (GE Medical Systems, Madison, Wisconsin, United States; software version 13.60.033). BMD was measured at the total hip and lumbar spine (L2 to L4), and T-scores were generated from the manufacturer’s reference database of sex-matched healthy young adults. Only hip T-scores were used for analysis. At the proximal femur, measurements from the neck, trochanteric, and intertrochanteric regions were combined to yield a total hip value. Algorithm development The framework was designed to answer a direct clinical question: given a patient’s current T-score trajectory, at what age will they cross into the osteoporotic range (T-score ≤ − 2.5) or, conversely, recover from it? Patient age and corresponding hip T-scores were processed using R version 4.5.1 (R Foundation for Statistical Computing, Vienna, Austria). Graphical outputs were generated using Python version 3.12 (Python Software Foundation) with the matplotlib library version 3.x. Decimal ages were derived for each scan from the scan date and the first-scan age to maintain numerical precision. Linear regression was selected as the primary modelling approach for three reasons. First, it produces directly interpretable parameters (slope and intercept) that map to measurable physiological quantities and satisfy the explainability requirements emphasised by Rudin and Lipton for high-stakes clinical decision support [10, 11]. Second, with typically two to five scan points per patient in osteoporosis surveillance, higher-order polynomial fits are prone to overfitting and produce unstable extrapolations beyond the observed data range. Third, a deterministic linear baseline provides a transparent reference against which future non-linear or hybrid machine-learning extensions can be benchmarked. The practical implications of this linearity assumption are examined in the Discussion. Handling of scan intervals, missing data, and numerical safeguards Scans were included as recorded in the clinical database. Irregular inter-scan intervals were treated as feasible inputs given the retrospective nature of routine surveillance. Missing T-scores resulted in exclusion of that scan but not the patient, provided at least two valid scans remained. Duplicate same-day scans with identical values were deduplicated. Two numerical safeguards were applied to onset-age predictions: trajectories with slope below 0.01 T-score units per year were classified as insufficient-slope and excluded from threshold-crossing predictions, since inversion of near-zero slopes yields uninterpretable extrapolations, and predicted threshold-crossing ages were required to fall within 0 to 150 years, with out-of-range values classified accordingly. Two-point model (TTOc; see Fig. 1) Fig. 1Workflow for the two-point (TTOc and TEOc) and regression-based (TTOt) algorithms. This schematic illustrates how serial age and T-score pairs are used to calculate slope, intercept, and the predicted threshold-crossing age at T-score − 2.5. The two-point model uses consecutive scan pairs, while the regression model fits a least-squares trajectory across all available scans Workflow for the two-point (TTOc and TEOc) and regression-based (TTOt) algorithms. This schematic illustrates how serial age and T-score pairs are used to calculate slope, intercept, and the predicted threshold-crossing age at T-score − 2.5. The two-point model uses consecutive scan pairs, while the regression model fits a least-squares trajectory across all available scans For each pair of consecutive scans, a straight line was fitted between two coordinates (x₁, y₁) and (x₂, y₂), where x is decimal age in years and y is the hip T-score. The slope was calculated as:\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:m=\frac{{y}_{2}-{y}_{1}}{{x}_{2}-{x}_{1}}$$\end{document} The line equation was then:\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:y=mx+C\hspace{1em}\mathrm{where\:}C={y}_{1}-m{x}_{1}$$\end{document} Solving for y = − 2.5 yielded the predicted age of entry into osteoporosis:\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:x=\frac{-2.5-C}{m}$$\end{document} If the patient was already osteoporotic (y ≤ − 2.5) and the slope was positive, the same calculation estimated the age at which the trajectory would exit the osteoporotic range (TEOc). Non-generating cases The algorithm computes slope and intercept for every scan pair but only reports a threshold-crossing age when the trajectory direction matches the clinical question (TTO for non-osteoporotic worsening, TEO for osteoporotic improving). Other combinations are classified as non-generating and no predicted age is reported, because mathematical extrapolation in these cases either describes a past crossing rather than a future one, or projects a threshold the patient is moving away from. Multipoint regression model (TTOt; see Fig. 1) For patients with three or more scans, a least-squares regression line was fitted through all available data points for that patient:\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:m=\frac{\sum\:\left({x}_{i}-\stackrel{-}{x}\right)\left({y}_{i}-\stackrel{-}{y}\right)}{\sum\:{\left({x}_{i}-\stackrel{-}{x}\right)}^{2}}\mathrm{\:\:where\:\:}C=\stackrel{-}{y}-m\stackrel{-}{x}$$\end{document} where xi is the patient’s decimal age at the i-th scan, yi is the hip T-score at the i-th scan, x̄ is the arithmetic mean of the patient’s decimal ages across all included scans, ȳ is the arithmetic mean of the patient’s T-scores across all included scans, and summation is taken from i = 1 to n (the number of valid scans). The threshold-crossing age was obtained as\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:x=\frac{-2.5-C}{m}$$\end{document} Uncertainty quantification for TTOt A 95% prediction interval was computed for the TTOt threshold-crossing age using inverse prediction from the per-patient regression. The standard error of the inverse prediction was calculated as:\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\eqalign{ SE\left( {\hat x} \right) = & \left( {\hat \sigma /\left| m \right|} \right) \cr & \times \surd \left( {1 + 1/n + \left( {\hat x - \bar x} \right)/\Sigma \left( {{x_i} - \bar x} \right)} \right) \cr} $$\end{document} where σ̂ is the residual standard deviation of the regression fit. The 95% prediction interval was \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:\hat x\:\pm\:\:t\_(n-2,\:0.975)\:\times\:\:SE(\hat x)$$\end{document}, where t denotes the Student’s t-critical value. Intervals were computed only for patients with three or more scans, since prediction intervals from two observations alone are not statistically meaningful. Evaluation strategy Three complementary validation analyses were performed. Onset-age validation For patients who started non-osteoporotic and crossed into the osteoporotic range during follow-up, predicted onset ages from TTOc (computed from the last pre-crossing scan pair) and TTOt (regression through all pre-crossing scans) were compared against the observed age at first osteoporotic DEXA reading. Agreement was quantified using mean absolute error (MAE), mean error (bias), and Bland-Altman limits of agreement. Prospective T-score validation For each patient with three or more scans, scans 1 to N − 1 were used to fit the model and predict the T-score at scan N. For TTOc, the last available pair was used. For TTOt, regression through all preceding scans was used. Predicted T-scores were compared to observed T-scores, and MAE and mean error were computed across the full validation set. Scan-count stability analysis For patients in the 5-or-more-scan stratum, TTOt was repeatedly computed using the first 2, 3, 4, and 5 scans progressively. Standard deviation of TTOt estimates and mean prediction interval width were tabulated at each scan count. Full pseudocode for TTOc, TTOt, and the prediction-interval computation is provided in Supplementary File 1. Results Cohort characteristics The cohort comprised 200 patients drawn from an eligible pool of 23,391 patients meeting inclusion criteria. Baseline characteristics are presented in Table 1. Patients were predominantly female (178, 89.4%), with a mean age at first scan of 58.3 years (SD 10.0, range 23 to 87). Mean body mass index was 25.2 kg/m² (SD 4.4). At baseline, 19 patients (9.5%) were already osteoporotic. Table 1Baseline characteristics of the study cohort (n = 200)CharacteristicValuePatients, n200Female, n (%)178 (89.4)Age at first scan, mean (SD), years58.3 (10.0)Age at first scan, range, years23 to 87BMI, mean (SD), kg/m²25.2 (4.4)Osteoporotic at baseline, n (%)19 (9.5)On osteoporosis pharmacotherapy at baseline, n (%)10 (5.0)Scan acquisition windowNovember 1999 to November 2024Patients with 2 scans, n50Patients with 3 scans, n50Patients with 4 scans, n50Patients with ≥ 5 scans, n (first 5 used for analysis)50 Table 1Baseline characteristics of the study cohort (n = 200)CharacteristicValuePatients, n200Female, n (%)178 (89.4)Age at first scan, mean (SD), years58.3 (10.0)Age at first scan, range, years23 to 87BMI, mean (SD), kg/m²25.2 (4.4)Osteoporotic at baseline, n (%)19 (9.5)On osteoporosis pharmacotherapy at baseline, n (%)10 (5.0)Scan acquisition windowNovember 1999 to November 2024Patients with 2 scans, n50Patients with 3 scans, n50Patients with 4 scans, n50Patients with ≥ 5 scans, n (first 5 used for analysis)50 Baseline characteristics of the study cohort (n = 200) Algorithm output Across 200 patients, the two-point algorithm generated 500 scan-pair predictions. Of these, 248 pairs produced a TTO prediction (non-osteoporotic patients with declining slope approaching the threshold), 34 produced a TEO prediction (osteoporotic patients with improving slope), 141 were classified as non-generating (slope direction away from threshold), 53 had insufficient slope magnitude for reliable inversion, and 21 produced out-of-range predicted ages. The regression algorithm (TTOt) produced valid threshold-crossing predictions for 163 of 200 patients (81.5%), with 95% prediction intervals available for 120 patients with three or more scans. Observable events Of the 181 patients who were non-osteoporotic at baseline, 19 (10.5%) experienced an observable crossing into the osteoporotic range during follow-up. Of the 19 patients who were osteoporotic at baseline, 4 (21.1%) experienced an observable recovery to T-score above − 2.5. The remaining 177 patients did not cross the diagnostic threshold during the observation window. The algorithm output reflected this clinical heterogeneity.Among the 162 non-osteoporotic baseline patients who did not cross the threshold during follow-up, the majority were classified by TTOc either as non-generating (slope direction away from threshold) or as producing predicted threshold-crossing ages beyond the observation window. This pattern indicates that the algorithm distinguishes patients on declining trajectories from those on stable or improving trajectories rather than producing threshold-crossing predictions indiscriminately. Onset-age validation Of the 19 patients with observed osteoporosis onset, 8 (42%) had sufficient pre-crossing data for algorithmic prediction; 8 had crossed at their second scan (leaving only one pre-crossing scan, insufficient for either algorithm); 2 had trajectories with insufficient slope or non-generating directions before crossing; and 1 had a positive pre-crossing slope reflecting a non-linear biological pattern. For the 8 patients with valid predictions, TTOc predicted observed onset age with a mean absolute error of 5.67 years (mean bias + 1.42 years, standard deviation of errors 8.87). The Bland-Altman 95% limits of agreement spanned − 16.0 to + 18.8 years. TTOt predicted observed onset age with a mean absolute error of 7.33 years (mean bias + 4.73 years, standard deviation 9.66). The Bland-Altman 95% limits of agreement spanned − 14.2 to + 23.7 years. Both algorithms demonstrated a small positive bias in onset prediction, tending to predict onset slightly later than observed. The wider prediction intervals for TTOt reflect the mathematical amplification of slope uncertainty during inverse prediction when only two or three pre-crossing scans are available. Prospective T-score validation For the 150 patients with three or more scans, 300 prospective T-score predictions were generated. TTOt produced lower error than TTOc on this task. The regression approach achieved a mean absolute error of 0.385 T-score units (mean bias + 0.004, SD 0.652), whereas TTOc achieved a mean absolute error of 0.533 T-score units (mean bias − 0.107, SD 1.968). The TTOc standard deviation reflects occasional large prediction errors driven by noise in the most recent scan pair, whereas TTOt’s smoothing across all available scans produced consistently closer predictions. For context, the intrinsic scan-to-scan measurement precision of hip DEXA is approximately 0.10 to 0.15 T-score units. TTOt’s prediction error of 0.385 is roughly three to four times this baseline. In practical terms, the model’s predictive uncertainty is comparable in scale to the measurement uncertainty that any single DEXA scan carries, rather than being orders of magnitude worse, as is common with extrapolation-based models. This strong short-horizon performance reflects the relative stability of BMD trajectories over typical surveillance intervals rather than any sophistication in the model, and it should not be extrapolated to long-range onset prediction, where the slope-inversion step amplifies uncertainty substantially. Scan-count stability analysis Prediction intervals for TTOt narrowed substantially with each additional scan (Table 2). Mean prediction interval width for the 5-scan stratum decreased from 654 years using the first 3 scans only, to 202 years using the first 4 scans, to 69.5 years using all 5 scans. Mean predicted TTOt values remained centred around 72 to 77 years across all scan counts, suggesting that the central estimate is relatively stable while the uncertainty around it is highly sensitive to data quantity. Table 2Algorithm output and validation metricsAlgorithm output (n = 200)MetricValueTTOc scan-pair predictions computed500TTOc generating TTO (valid predictions)248TTOc generating TEO (valid predictions)34TTOc non-generating (slope direction)141TTOc insufficient slope magnitude53TTOc out-of-range predicted age21TTOt valid threshold-crossing predictions163 / 200TTOt 95% prediction intervals computed120Onset-age validation (observed crossings with sufficient pre-crossing data,n = 8) Metric TTOc TTOt N with valid prediction88Mean absolute error, years5.677.33Mean bias (predicted − observed), years+ 1.42+ 4.73Standard deviation of errors, years8.879.66Bland-Altman 95% limits of agreement, years−16.0 to + 18.8−14.2 to + 23.7 Prospective T-score validation (300 predictions in 150 patients) Metric TTOc TTOt Mean absolute error, T-score units0.5330.385Mean bias, T-score units−0.107+ 0.004Standard deviation, T-score units1.9680.652Scan-count stability analysis (TTOt, 5-scan stratum) Scans used N valid Mean TTOt, yearsSD of TTOt, yearsMean 95% PI width, years24067.522.6not applicable34277.227.265444272.633.320253673.229.069.5 Table 2Algorithm output and validation metricsAlgorithm output (n = 200)MetricValueTTOc scan-pair predictions computed500TTOc generating TTO (valid predictions)248TTOc generating TEO (valid predictions)34TTOc non-generating (slope direction)141TTOc insufficient slope magnitude53TTOc out-of-range predicted age21TTOt valid threshold-crossing predictions163 / 200TTOt 95% prediction intervals computed120Onset-age validation (observed crossings with sufficient pre-crossing data,n = 8) Metric TTOc TTOt N with valid prediction88Mean absolute error, years5.677.33Mean bias (predicted − observed), years+ 1.42+ 4.73Standard deviation of errors, years8.879.66Bland-Altman 95% limits of agreement, years−16.0 to + 18.8−14.2 to + 23.7 Prospective T-score validation (300 predictions in 150 patients) Metric TTOc TTOt Mean absolute error, T-score units0.5330.385Mean bias, T-score units−0.107+ 0.004Standard deviation, T-score units1.9680.652Scan-count stability analysis (TTOt, 5-scan stratum) Scans used N valid Mean TTOt, yearsSD of TTOt, yearsMean 95% PI width, years24067.522.6not applicable34277.227.265444272.633.320253673.229.069.5 Algorithm output and validation metrics The number of patients producing valid TTOt estimates at each scan count varies because trajectories with insufficient slope or out-of-range predicted ages were excluded under the numerical safeguards described in Methods. Recovery case Among the 4 patients with observable recovery to non-osteoporotic T-scores, one had sufficient pre-recovery data for algorithmic prediction. For this patient (stratum: ≥5 scans), observed recovery age was 69 years; both TTOc and TTOt predicted recovery at 66.8 years, representing a 2.2-year early prediction. This single case is reported as an illustrative observation rather than statistical evidence of recovery-prediction performance. Worked examples Two representative cases illustrate the complementary behaviour of the algorithms (Fig. 2). In the first subject (Fig. 2a, a female patient with five scans spanning approximately 18 years and monotonically declining T-scores), the observed onset age was 55.0 years. TTOc, fitted from the last pre-crossing scan pair, predicted onset at 51.0 years (error − 4.0 years). TTOt, using regression through all pre-crossing scans, predicted onset at 53.1 years (error − 1.9 years). Both predictions were within a clinically useful range. Fig. 2Trajectory plots of two representative patients with observed osteoporosis onset from the n = 200 cohort. (a) A female patient with five hip DEXA scans spanning approximately 18 years (from age 43 to age 61), showing monotonic T-score decline and entry into the osteoporotic range at age 55.0 years. TTOc (fitted from the last pre-crossing scan pair) predicted onset at 51.0 years; TTOt (regression through all three pre-crossing scans) predicted onset at 53.1 years. Both algorithms slightly underestimated onset age in this case, a bias direction consistent with the cohort-level finding. (b) A female patient with five hip DEXA scans spanning approximately 11 years (from age 63 to age 74), showing non-uniform trajectory and entry into the osteoporotic range at age 71.1 years. TTOc predicted onset at 66.3 years, while TTOt predicted onset at 85.4 years. The divergence arises because the last pre-crossing scan pair is closely spaced with a steep local slope, whereas the regression smooths across all three pre-crossing scans including an early near-flat segment. This case illustrates how the two algorithms capture different aspects of the same trajectory. Open circles: pre-crossing scans used for prediction. Filled black circle: observed crossing scan (T ≤ − 2.5). Grey circle: post-crossing scan (shown for context; not used in prediction). Blue segment and triangle: TTOc prediction. Orange segment and triangle: TTOt prediction. Dashed horizontal line: osteoporosis threshold at T = − 2.5 Trajectory plots of two representative patients with observed osteoporosis onset from the n = 200 cohort. (a) A female patient with five hip DEXA scans spanning approximately 18 years (from age 43 to age 61), showing monotonic T-score decline and entry into the osteoporotic range at age 55.0 years. TTOc (fitted from the last pre-crossing scan pair) predicted onset at 51.0 years; TTOt (regression through all three pre-crossing scans) predicted onset at 53.1 years. Both algorithms slightly underestimated onset age in this case, a bias direction consistent with the cohort-level finding. (b) A female patient with five hip DEXA scans spanning approximately 11 years (from age 63 to age 74), showing non-uniform trajectory and entry into the osteoporotic range at age 71.1 years. TTOc predicted onset at 66.3 years, while TTOt predicted onset at 85.4 years. The divergence arises because the last pre-crossing scan pair is closely spaced with a steep local slope, whereas the regression smooths across all three pre-crossing scans including an early near-flat segment. This case illustrates how the two algorithms capture different aspects of the same trajectory. Open circles: pre-crossing scans used for prediction. Filled black circle: observed crossing scan (T ≤ − 2.5). Grey circle: post-crossing scan (shown for context; not used in prediction). Blue segment and triangle: TTOc prediction. Orange segment and triangle: TTOt prediction. Dashed horizontal line: osteoporosis threshold at T = − 2.5 In the second subject (Fig. 2b, a female patient with five scans showing more complex trajectory behaviour), the observed onset age was 71.1 years. TTOc predicted onset at 66.3 years (error − 4.7 years), while TTOt predicted onset at 85.4 years (error + 14.3 years). The divergence between the two algorithms in this case illustrates how the two approaches capture different aspects of trajectory: TTOc’s focus on the most recent pair detected the acceleration preceding the crossing, while TTOt’s averaging across the full pre-crossing history diluted this signal. Discussion This proof-of-concept shows that simple, deterministic mathematical algorithms can convert serial DEXA T-scores into interpretable age-based trajectory predictions. Adherence to osteoporosis therapy remains a major challenge, and conventional reporting often provides limited guidance. Real world persistence with oral bisphosphonates varies widely, with medication possession ratios that are frequently low [6]. Systematic reviews suggest that multicomponent interventions involving education, counselling, and active patient involvement are more likely to improve adherence than single component approaches [7]. Evidence also indicates that personalised, numeric communication of osteoporosis risk may improve patient understanding and intent to treat when compared with qualitative descriptions [8]. In keeping with findings from the RICO study [8], our model may help by translating abstract T scores into personalised, age-based predictions that patients can interpret more easily. Whether such translations meaningfully improve clinical communication, risk assessment, or shared decision-making is an empirical question that a single-centre retrospective analysis cannot answer and that future validation studies should address directly. Complementary strengths of TTOc and TTOt An unexpected but instructive finding is that TTOc and TTOt are complementary rather than redundant. TTOc, by using only the most recent scan pair, captures acceleration in decline and produced the lower onset-age error in this validation subset. TTOt, by averaging across all available scans, yields more stable estimates of short-term trajectory and predicts next-scan T-scores more precisely. No single linear formulation is therefore optimal for all aspects of trajectory reporting. This pattern suggests that clinical tools built on this foundation may benefit from presenting multiple views of the same data, and it has direct implications for how future validation studies should be designed. Onset-age and trajectory prediction are distinct predictive tasks with distinct reliability characteristics, and validation against appropriately chosen endpoints should reflect this. The biological basis for this complementarity aligns with known patterns of bone mineral density change. Declines in T-score are often non-uniform over time, with accelerated loss around menopause, periods of immobilisation, or initiation of corticosteroids [9]. A two-point algorithm focused on recent scans naturally captures such acceleration, whereas a regression smoothed across many years may dilute it. Conversely, for predicting the short-term trajectory of T-score itself, the regression approach benefits from the statistical strength of using all available measurements. Stability and minimum-scan requirements The scan-count stability analysis establishes a quantitative minimum-data requirement for reliable interval estimation. TTOt prediction intervals narrowed dramatically with additional scans, from a mean width of 654 years at three scans to 69.5 years at five scans. This represents a near-tenfold reduction in uncertainty between the third and fifth scan and reflects the extreme sensitivity of slope-inverted predictions to small-sample regression variance. Two practical implications follow. First, the finding offers an empirical basis for reporting conventions. TTOt should be presented as a precise prediction only when four or more scans are available, with TTOc preferred as a provisional indicator below this threshold. Second, it provides a quantitative foundation for future evaluation of serial imaging frequency. The steep non-linearity of the scan-count benefit curve suggests that the clinical value of each additional scan is substantial within the three-to-five-scan range, an observation that prospective cohort studies could test against surveillance outcomes. Interpretability and its clinical stakes Beyond numerical accuracy, the usability of any predictive method depends on its transparency. Informatics literature has highlighted that opaque predictive systems reduce user trust, complicate validation, and obscure failure modes. Rudin has argued that high-stakes clinical decisions should rely on inherently interpretable models rather than post hoc explanation, since transparency is essential for safety and auditability [10]. Lipton has similarly noted that parameter-level interpretability allows end users to understand and challenge predictions when required [11]. Each parameter of the proposed framework (slope, intercept, threshold-crossing age) is derived directly from measured DEXA T-scores, which are themselves validated clinical indicators of bone mineral density. The entire computation fits on a single page and can be reproduced by a clinician with a calculator. This structural transparency has a specific informatics payoff. A clinician who receives a prediction of “TTOc 77 years” for a 72-year-old patient can inspect the two T-scores, the two ages, and the slope that produced the estimate. An algorithmic decision that disagrees with clinical intuition can be interrogated, challenged, or overridden with the reasoning intact. Opaque predictive systems, whether classical machine learning or deep learning, do not permit this kind of interrogation without additional explanatory infrastructure. In settings where deployed AI is expected to meet emerging standards for explainability, a deterministic baseline of this kind is more readily inspectable than opaque alternatives. Implications for future development These findings point toward three directions that future work should pursue before the framework could inform clinical practice. First, whether embedding TTOc values into DEXA reports improves clinical communication is directly testable. A trial, assigning clinicians to standard or TTOc-augmented reports could measure whether the additional information changes preventive discussions, referral patterns, or patient understanding. The algorithm’s transparency makes this an unusually tractable intervention study compared with opaque predictive models, because clinicians and patients can inspect the reasoning behind each prediction rather than reacting to a black-box output. Second, TTOt prediction intervals offer a natural vehicle for communicating uncertainty, but their usability remains untested. At five scans, intervals averaged approximately 70 years wide and whether clinicians and patients find such ranges useful, misleading, or uninterpretable is itself an empirical question. Work on numeric risk communication in the decision-science literature provides a methodological starting point. Third, whether denser monitoring schedules translate into earlier detection and better outcomes is an open question for future prospective work. Assumption of linearity and alternative modelling approaches The linearity assumption is the most important methodological simplification in this work. Bone loss in real patients is often non-linear, with accelerated decline during the perimenopausal transition, plateaus or reversals in response to pharmacologic therapy, and age-related slowing in very late life. Linear extrapolation can therefore misestimate the predicted age of threshold crossing in patients whose biology is actively changing. Several alternative modelling approaches warrant exploration in larger datasets. Piecewise linear regression with a breakpoint at menopause could accommodate accelerated post-menopausal decline. Segmented regression triggered by treatment initiation could capture pharmacologic effects on slope. Non-linear parametric models such as exponential-decay functions may fit untreated long-term decline better than straight lines. Generalised additive models and Gaussian process regression offer flexible trajectory smoothing while retaining interpretability. Mixed-effects models with patient-level random slopes and intercepts would be the natural extension for population-level trajectory estimation with individual-level precision. Confounders, treatment, and adjustment This proof-of-concept did not adjust for clinical confounders that are known to modify BMD trajectories, including sex, age at first scan, BMI, menopausal status, corticosteroid exposure, and osteoporosis pharmacotherapy. Bisphosphonates, denosumab, and teriparatide can substantially alter the slope of T-score change, and failure to account for treatment initiation mid-trajectory may produce misleading predictions in either direction. Our approach treats the observed T-score sequence as the input, without decomposing it into pre- and post-treatment segments. Future extensions should incorporate treatment start and stop dates as covariates, either through segmented regression or through mixed-effects models with treatment as a time-varying predictor. External validation and future work External validation in larger, demographically diverse cohorts is the essential next step. Candidate datasets include national DEXA registries in Ireland and other European countries, the UK Biobank subset with serial DEXA measurements, and collaborative multi-centre cohorts assembled through the International Osteoporosis Foundation. A concrete validation roadmap would (a) replicate both TTOc and TTOt in an independent cohort using the same pre-specified stratified sampling approach, (b) compare observed versus predicted agreement using the metrics reported here, (c) stratify performance by sex, BMI, treatment status, and inter-scan interval, and (d) benchmark the deterministic approach against mixed-effects and Bayesian alternatives. Limitations Several limitations frame the interpretation of these findings. First, the onset-age validation was conducted on 8 patients whose pre-crossing data permitted algorithmic prediction, a natural consequence of the rarity of observed threshold crossings in typical DEXA surveillance cohorts and of the minimum data requirements of each algorithm. Second, the Bland-Altman limits of agreement were wide at the individual patient level, ranging from − 16.0 to + 18.8 years for TTOc and from − 14.2 to + 23.7 years for TTOt, indicating that point estimates should be interpreted alongside their uncertainty ranges and not as precise predictions. Third, both models demonstrated a small positive bias in onset prediction, tending to predict onset later than observed, which has implications for clinical safety and should be accounted for when considering deployment. Fourth, the recovery-age validation rested on a single patient with evaluable pre-recovery data and must be regarded as a case observation. Fifth, the study was conducted in a single institutional cohort with predominantly female representation (89.4%), limiting generalisability across sex and geography. Sixth, DEXA reproducibility error of approximately 1 to 2% influences slope estimation, particularly in two-point calculations with closely spaced scans. Seventh, the linear model does not capture non-linear biological patterns such as accelerated perimenopausal decline or treatment-induced plateaus. Conclusions The framework presented here is best understood as a transparent methodological foundation, a deterministic baseline against which more sophisticated models can be benchmarked and onto which clinically meaningful extensions can be built. Each parameter corresponds directly to a measurable property of bone physiology, satisfying the interpretability requirements that distinguish trustworthy clinical decision support from opaque prediction. The complementary behaviour of the two-point and regression algorithms is itself a useful finding, that no single linear formulation is optimal for all aspects of trajectory reporting, and future clinical tools may benefit from presenting multiple views of the same data. What remains to be established is whether trajectory-based reporting, however mathematically sound, actually changes clinical outcomes. That question lies beyond the scope of any single-centre retrospective analysis and requires prospective evaluation against meaningful endpoints. The present work offers a concrete starting point for such evaluation: a transparent algorithm, validated against observed outcomes, with explicitly quantified uncertainty and clearly bounded claims. For bone health informatics, this combination of properties may be a useful complement to higher-performing but less inspectable models. Electronic Supplementary Material Below is the link to the electronic supplementary material. Supplementary Material 1: Pseudocode for TTOc, TTOt, and prediction-interval computation Supplementary Material 1: Pseudocode for TTOc, TTOt, and prediction-interval computation Supplementary Material 2. Ethics letter from the UL Hospital Group Research Ethics Committee Supplementary Material 2. Ethics letter from the UL Hospital Group Research Ethics Committee Supplementary Material 3: Per-patient TTOc and TTOt predictions for the full cohort (n = 200), with 95% prediction intervals and observed onset age where applicable Supplementary Material 3: Per-patient TTOc and TTOt predictions for the full cohort (n = 200), with 95% prediction intervals and observed onset age where applicable