Intelligent vulnerability assessment system for college students mental health based on few-sample reinforcement learning approach Xiaobo Huang, Jiayue Ji, Shi Cheng Abstract AbstractIntroductionCollege students’ psychological vulnerability assessment is constrained by data scarcity, high labeling cost, privacy sensitivity, and small, imbalanced samples, and existing frameworks typically treat vulnerability as a static judgment that cannot track how a student’s condition changes over time. This study examined whether a few-sample hierarchical grading model combined with meta-reinforcement-learning-based sequential decision-making could address both problems jointly.MethodsA few-sample hierarchical grading model, combining meta-learning and transfer regularization with a prototype-based embedding and ordinal threshold assignment, quantified static risk from a handful of labeled examples per grade, while a sequential risk-identification method formulated dynamic assessment as a Markov decision process solved by a two-stage meta-reinforcement-learning policy. Case studies used a longitudinal, class-imbalanced cohort simulation calibrated to the documented statistics of a mobile-sensing mental-health study (no real participant data were used) and compared the proposed models against threshold-rule, supervised, fine-tuned, few-shot, and non-meta reinforcement-learning baselines, with sensitivity analyses on support-set size, reward weights, discount factor, and regularization weight, plus a regime-switching robustness check.ResultsThe proposed grading model attained a macro-F1 of 0.847 with ten labeled examples per grade, over twenty percentage points above a conventional deep-learning classifier, though its edge over other episodic few-shot architectures was small and inconsistent, indicating episodic meta-learning as the main source of this gain. The meta-reinforcement-learning policy achieved a higher alert precision-recall area (0.514 vs. 0.380) than a non-meta baseline and avoided its degenerate always-escalate behavior, but neither the main study nor the regime-switching check showed the composite reward lengthening lead time versus a simpler accuracy-only reward, which matched its lead time at a much lower escalation rate. Sensitivity analyses showed the reward-weight ratio and discount factor shaping the early-identification versus intervention-cost trade-off.ConclusionThese findings suggest that combining few-sample meta-learning with metareinforcement-learning-based sequential decision-making can improve vulnerability assessment under scarce, evolving data relative to conventional static and non-meta-adapted approaches, with an accuracy-only reward recommended as the practical default. The case studies relied on a calibrated simulation rather than raw longitudinal records, so prospective validation on real campus data under appropriate privacy safeguards is needed before responsible deployment.Keywords: college student mental health, few-shot learning, meta-learning, ordinal classification, psychological vulnerability assessment, reinforcement learning, sequential decision-making Introduction Nowadays, mental health problem of college students has become a more and more urgent global concern, and elevated rates of depressive symptoms, anxiety, and suicidal ideation have been reported across diverse student populations, which are also shown to cluster with broader health risk behaviors (). Besides, public health disruptions, such as the prolonged campus quarantines during the COVID-19 pandemic, have further increased the psychological strain on students and exposed the fragility of existing campus support mechanism (). Not only individual well-being is affected, unresolved psychological vulnerability also carries a real institutional consequence: administrative evidence has linked persistent student mental health difficulties with elevated dropout from higher education (), and delayed identification will bring a real cost to the institution. Because of this, universities and policymakers have gradually moved from ad hoc counseling arrangement toward a more formalized, institution-wide mental health policy that is aimed at systematizing the early identification and intervention (). In recent years, this policy shift has also started to meet with technological innovation, since artificial-intelligence-enabled tools are more and more explored as a way to extend the reach and responsiveness of campus mental health support (). Under this background, it becomes a pressing and practical need for higher education institutions to develop an assessment approach that can identify the psychologically vulnerable students in an accurate, efficient, and large-scale way. Standardized psychological scales and questionnaires are still the most used instrument by which universities carry out large-scale mental health screening, and tools such as the General Health Questionnaire are routinely used to classify students by their risk level (). However, more and more psychometric studies have found that this kind of instrument cannot always generalize well across different subpopulations, since reliability and measurement invariance analyses in college student athletes have shown that questionnaire batteries which are validated on general population may not hold consistently across different student groups (). Besides, symptom-network analyses of large student cohorts have further argued that the total-score paradigm which most scales are based on will obscure the specific symptoms that most strongly drive an individual’s overall risk, so the diagnostic precision of aggregate scores becomes limited (). What makes this problem worse is that evidence from universal self-report screening programs shows that scale-based assessment can also bring some unintended harm, including false positives and screening-related distress, when it is applied to entire student populations without careful consideration (). These limitations have made researchers try to shorten and personalize the assessment scales in order to reduce the burden on respondents while the diagnostic value is preserved, but such adaptation still relies on the same static, single-timepoint measurement logic, and it does little to solve the underlying cost of large-scale data collection and expert labeling (). In order to go beyond the limitation of self-report instrument, many works have already applied machine learning technique to predict the psychological distress and risk directly from survey, cohort, or longitudinal data, and it is shown that data-driven models can capture the risk patterns which the static scale score alone usually cannot catch (, ). Electronic-health-record-based models have also extended this line of work into clinical setting, showing that mental health crisis can be anticipated in advance from the patient data that is routinely collected, instead of relying only on manual chart review (), and at the same time, cross-cultural machine learning studies have tried to find out which risk factor can most reliably predict the adolescent mental health outcome across different population (). Specifically in higher education, deep learning approaches have been proposed for detecting the stress severity from physiological pulse signal and for automatically identifying depressive symptom from student’s online behavioral trace, so that the assessment can move away from self-assessment scale and toward the continuously collected behavioral evidence (, ). Multimodal deep learning framework which integrates biological, lifestyle, text, speech, and physiological signal has also pushed the prediction performance further, and this shows the potential of combining heterogeneous data stream for early mental health risk detection (, ). Even so, a recent and comprehensive review of this literature concludes that the reported gain of machine learning and deep learning model is, in almost every case, depending on the access to large and well-labeled training dataset, and this assumption is hard to satisfy in campus mental health context, because data collection there is constrained by privacy sensitivity, stigma, and limited institutional resource (). This dependency on data has made researchers turn to few-shot learning, meta-learning, and other related low-resource paradigm, which are designed on purpose to reduce the reliance on large annotated corpus. Lightweight, instruction-tuned language model and prompt-based learning framework have been proposed to detect depression from limited labeled data, and it is reported that they can achieve competitive accuracy while the computational and annotation cost is reduced a lot (, ). Based on this direction, few-shot and zero-shot prompting strategy have also been applied to identify depression and anxiety symptom in patient communication and to distinguish depression subtype from social media text, and several studies have compared few-shot prompting with conventional fine-tuning directly, reporting that a carefully designed prompt can match or even approach the fine-tuned performance under scarce-label condition (, ). Related work has used large-language-model-based data augmentation to relieve the scarcity of labeled suicidal ideation text, and multitask learning over pretrained audio and text embedding has been used to predict depression severity and suicide risk together from small clinical sample (, ). Not only for text application, few-shot learning has also been proven effective for emotion recognition from spontaneous speech and for decoding the mental state from fMRI signal under limited-data condition, and a broader survey of few-shot learning for biomedical time series confirms that this paradigm is more and more viewed as a possible way to build clinically meaningful model when large labeled dataset is not available (–). Even so, almost all of these efforts still treat the psychological assessment as a one-shot, static classification problem, which means predicting one single risk label from a fixed input, and they do not go further to model how an individual’s psychological vulnerability changes and needs to be tracked as time goes on. Reinforcement learning can offer a complementary paradigm here, because it is designed for sequential decision-making instead of single-shot prediction, and its application on health-related risk assessment has started to get attention recently. A systematic review of reinforcement learning in critical care has concluded that this technique is very suitable for the problem which needs repeated, state-dependent decision, although the review also points out that the clinical adoption is still at an early stage (). Disease progression has already been modeled as a sequential reasoning process over a medical knowledge graph, in which an agent takes a series of interpretable step to arrive at the risk prediction, and this shows how sequential decision formulation can support dynamic, explainable risk trajectory rather than a static label (). At the same time, adaptive trial design which adjusts the intervention strategy step by step based on the accumulating evidence, and Bayesian sequential monitoring scheme which keeps updating risk estimate as new surveillance data arrives, both demonstrate that it is valuable to treat the health-related risk identification as an evolving process instead of a one-time assessment (, ). Outside the health domain, reinforcement learning has also been used for identifying and predicting the dynamically evolving risk in cost-sensitive credit decision and in corporate low-carbon transition management, and both of them model the risk identification explicitly as a sequential decision problem, with an agent that keeps updating its assessment policy through interacting with the incoming data (, ). However, application of reinforcement learning specifically on psychological vulnerability assessment is still very scarce, and none of the studies reviewed above has combined the sequential decision-making with the low-resource learning capability that is needed to work under the data constraint of college mental health setting. If we put all the reviewed literature together, three limitations can be found that current approach to college student mental health assessment has not solved yet. First, traditional scale- and questionnaire-based assessment is still constrained by difficult data acquisition, high labeling cost, privacy sensitivity, and small and imbalanced sample, and it also suffers from the reliability gap across different subpopulation and the risk of screening-related harm. Second, although machine learning and deep learning model can perform better than static scale, their reported gain depends heavily on large labeled dataset that is rarely available in campus mental health context, and even the newly emerging few-shot and meta-learning solution is mostly limited to one-shot, static classification, without being extended systematically to psychological vulnerability grading under real data scarcity. Third, almost all existing evaluation framework, no matter it is scale-based or data-driven, treats the psychological assessment as a single-point-in-time judgment, so the dynamic and individually different evolution of psychological vulnerability is left largely unaddressed, even though reinforcement-learning-based sequential decision-making has already shown its promise for dynamic risk identification in other health and non-health domain. In order to address these gaps, this paper proposes an intelligent vulnerability assessment system for college students’ mental health, which integrates few-shot learning with reinforcement-learning-based sequential decision-making. The main contribution of this paper can be summarized as follows: 1) a few-shot learning-based hierarchical assessment model is established for grading the college students’ psychological vulnerability, in which episodic meta-learning is the primary source of the model’s fewsample generalization advantage over a conventional, non-episodically-trained classifier, and transfer regularization anchored to a pretrained backbone is incorporated as a complementary refinement that is concentrated at the smallest support sizes, so the reliance on large-scale labeled data can be alleviated and the model’s generalization ability under few-shot condition can also be improved; 2) based on this hierarchical assessment model, a reinforcement-learning-based method for temporal identification of psychological risk is further proposed, which formulates the identification process as a sequential decision problem with the design of state, action, and reward, so that an agent can interact with the assessment environment and adjust its strategy dynamically, in order to track and predict the evolving trajectory of a student’s potential psychological risk; the empirical role of the reward’s cost and timeliness terms relative to an accuracy-only formulation, and the practical operating point this implies, are examined in detail in Section 5 and are found to favor a more conservative reward configuration than the full composite reward for most deployment settings. The rest of this paper is organized in the following way. Section 2 describes the problem background and formalizes the college student mental health vulnerability assessment problem, and it characterizes both the static grading task and the dynamic risk identification task that together motivate the proposed modeling framework. Section 3 presents the mathematical formulation of the few-shot hierarchical assessment model and the reinforcement-learning-based sequential risk identification method, together with the corresponding solution approach. Section 4 reports the simulation-based case studies which validate the effectiveness of the proposed system, and it also conducts sensitivity analysis with respect to the key parameters and sample conditions, reporting the resulting metric primarily in descriptive terms. Section 5 then interprets these results in relation to the operational trade-offs a deploying institution would face, the few-shot and sequential clinical-decision-making literature reviewed above, and the limitations of a simulation-based validation. Finally, Section 6 concludes the paper by summarizing the main finding and discussing the direction for future research. Problem description Multi-source data-driven formulation of the psychological vulnerability assessment problem The psychological vulnerability assessment problem addressed in this paper comes from a concrete institutional need: a university counseling office must decide, for each student who is enrolled, how psychologically at risk that student currently is, by using whatever information the institution can realistically collect. In practice, this information is neither abundant nor uniform. There are three broad categories of data that are typically available. The first category is the historical psychological scale record, which means the score or item-level response from the periodic mental health screening that a subset of students have completed at some point during their enrollment. The second category is the campus behavioral log, including the library and dormitory access pattern, academic performance trajectory, campus card consumption regularity, and online learning platform engagement, and all of them are passively generated as a by-product of ordinary student life. The third category is the peer- and counselor-reported qualitative note, such as the referral record from resident advisor or the brief case note from voluntary counseling session. The modeling framework in this section and in Section 3 is defined over all three sources; the case studies in Section 4, however, are conducted on a calibrated simulation environment that reproduces only the first two sources, because the publicly documented dataset used for calibration does not release peer- or counselor-reported qualitative note, and this scope restriction is stated explicitly where the case-study data are described (Section 4.1). None of these three sources is complete for the whole student body, none of them is collected on a fixed schedule, and only a small part of students have a verified psychological risk label at any given time, because obtaining such a label usually needs a clinical interview or a validated scale given by trained staff. So the resulting dataset is small in absolute size, sparsely and unevenly labeled among students, and skewed toward the majority of students who are not in visible distress, and this combination of property is exactly what makes conventional large-sample supervised learning hard to apply directly on this task. Besides the general problem of scarcity, each of the three data sources also brings its own specific obstacle. Historical scale record is the most directly interpretable one in psychological term, but it is only collected once or twice per academic year during the scheduled screening event, so the interval between two consecutive scale administration for the same student can be many months, and during this period the student’s condition may already change a lot without being recorded. Behavioral log, on the other hand, is continuously generated and therefore much more abundant in volume, but it is only a weak and indirect proxy of psychological state: a decline of library visit or dormitory check-in could reflect an emerging withdrawal and distress, but it could also just reflect an unrelated change of study habit, a part-time job, or simply a personal preference, and no behavioral signal alone can carry a verified label to tell which interpretation is correct. Peer- and counselor-reported note carries the richest qualitative signal about a student’s real condition, but it is also the scarcest one among the three sources, since it depends on a referral or a voluntary counseling visit that has already happened, and it is recorded in free-form language which must be coded into a usable feature before it can be combined with the other two sources. So the general problem formulation introduced next must not only handle a small total sample size, but also this kind of asymmetry in reliability, frequency, and interpretability among the three sources that together make up the feature vector x. To think about this situation in a way that does not depend on any particular modeling choice, it is useful to first write down the assessment task in its most general form. Let each student be characterized, at the moment of assessment, by a feature representation x which is obtained by putting the three data sources described above together into one single vector, and let y denote the psychological risk label that the assessment system is supposed to infer for that student. Then the general assessment problem can be stated as constructing a mapping from x to y that is accurate for the students whose true label is not known, by only using the limited set of students for whom a label happens to be available. This formulation leaves the internal structure of the mapping unspecified on purpose, because the purpose of Section 2 is only to characterize the problem before we commit to a model, and the internal structure will be given only in Section 3. What matters at this stage is that the mapping has to be learned from an evaluation set which is small compared with the diversity of the student population, and also that the feature representation x itself is heterogeneous, combining the numerical scale score, count-based behavioral statistic, and coarsely coded qualitative indicator all within the same vector. A large-capacity model that is naively applied cannot solve this situation, because the temptation to simply “collect more data and train a bigger network” is the default response in many other application domains. A high-capacity classifier that is trained end-to-end on the pooled labeled subset described above only has, in effect, as many independent training signals as there are labeled students, and in a typical campus counseling office this may be just a few hundred case spread unevenly across risk grade and academic cohort. Under this condition, a flexible model tends to memorize the idiosyncratic combination of behavioral and scale feature that is associated with the small number of observed high-risk case, instead of learning a decision boundary that can generalize to the next cohort of student, whose behavioral pattern, course load, and life circumstance are different from those in the training set. This kind of failure is exactly why the problem should be treated as a family of related few-sample task, rather than one large classification problem that is solved once and reused forever, and this is also the reason why the general formulation introduced next is left open on purpose about how the mapping from x to y is learned. Both the objective and the constraint of this general problem come from the practical circumstance of campus mental health work, not from any convenience of modeling. The objective is to keep the assessment accurate and, what is more important, to generalize well to the students whose data pattern has never been seen before, even though the number of labeled example that is available for training is small. This objective cannot be given up, because a counseling office cannot wait for assessment until a large labeled dataset has been accumulated, and the students who are currently at risk need to be identified with whatever data already exist now. The constraint, on the other hand, comes from three sources that were already mentioned in Section 1 and are now translated into property of the problem: the data privacy consideration limits how much additional psychological data can be collected beyond what is already routinely gathered, which caps the sample size that can be achieved; the practical need of timely intervention limits how long the system is allowed to wait to accumulate more labeled case before it produces a usable assessment; and the genuine individual difference means the same behavioral pattern, for example a period of reduced social activity, can correspond to very different risk level for different student, so a single global mapping which is estimated from too few example per subpopulation may become systematically miscalibrated for exactly the student it is supposed to protect. summarizes this general formulation as a three-stage pipeline, which connects the data layer, a feature aggregation stage that produces the unified representation x, and an assessment mapping stage that produces the output y, and the constraint of data scarcity, labeling cost, and class imbalance are explicitly marked as external pressure acting on this pipeline, instead of being internal modeling choice. The general framework shown in is on purpose left open about how the assessment mapping is learned and about whether the resulting assessment refers to one single point in time or to a condition that is evolving. These two open questions are exactly where the present research is different from a conventional supervised-learning treatment of the problem, and both of them are also the entry point for the two innovations introduced in this paper. If left unresolved, the general framework in would just become an ordinary classifier trained on whatever labeled data is available, which, as it is argued in Section 1, tends to overfit under the sample size that is realistically available in campus setting, and it can only produce a static, one-shot judgment which cannot track how a student’s condition changes after the first assessment. So Section 2.2 will now reformulate these two open questions explicitly, and each of them is translated into a change of the structure of the general problem, instead of being left as only a narrative motivation. Problem reformulation for few-sample hierarchical grading and sequential risk identification The first reformulation is about how the mapping from x to y can be learned when the labeled example is scarce. Instead of treating the assessment problem as one single classification task trained on one pooled dataset, this paper reformulates it as a distribution of related few-sample task. Each task is built by drawing, from the bigger pool of partially labeled student record, a small support set of labeled example together with a query set of example whose label needs to be inferred, so a task with C risk grade and K labeled example per grade is called a K-shot, C-way task. Framing the problem in this way changes what is actually being learned: instead of estimating one fixed decision boundary from one small labeled set, the system is now asked to learn how to construct a decision boundary quickly from whatever small labeled set it is given, by practicing this construction again and again across many such task sampled from a task distribution. This reformulation makes it possible to apply meta-learning and transfer learning to the problem in a meaningful way, since both technique are based on this task-distribution view rather than one single fixed training set, and this is also the structural embodiment of the first innovation claimed in Section 1. In a practical way, the task Tn described above does not need to be constructed artificially, because it can arise naturally from how campus mental health data already accumulates. Each academic term, each college inside the university, or each counseling cohort that goes through a screening cycle can be treated as producing one task, since the students within this kind of group share a common assessment period and a common pool of partially labeled record from which a support set and a query set can be drawn. Framing the historical data in this way means that the episodic task structure needed for meta-learning is obtained basically for free from the institution’s already existing organizational unit, instead of needing an artificial re-partitioning of the data that has no correspondence with how the university actually operates, and this observation is what makes the few-sample reformulation practically implementable, not just a theoretical convenience. A closely related, but different, structural change is about the nature of the label space itself. The general formulation in Section 2.1 left y unspecified about whether it is a nominal category or an ordered grade. In this paper, y is set explicitly to be an ordinal risk grade, for example ranging from low to severe vulnerability, so misclassifying a severely at-risk student as only low-risk is a qualitatively more serious mistake than misclassifying a moderate case as low, even though both of them are technically a classification error of one grade. To represent this ordering is a structural requirement, not just a modeling detail, because a plain multi-class classifier trained under few-sample condition has no mechanism to prefer the “close” error over the “far” error, and under scarce label the two kinds of error can be easily confused. So the reformulated problem requires that the decision function respects the ordering of the grade set, and that adjacent grade are separated by a graded decision structure instead of an arbitrary one; this requirement will be carried into the ordinal thresholding construction of Section 3.1, but it is already introduced here as a defining property of the assessment task, which makes it different from an unordered few-shot classification problem. A simpler alternative for the ordinal reformulation, that is, training several independent binary classifier for each pair of adjacent grade and combining their output, is not chosen here for a related reason. Under the sample size that is available for each grade in a campus setting, splitting the already limited support set into several separate binary comparisons would leave each individual classifier with even fewer effective example than the joint prototype-based representation described above, and this makes the data-scarcity constraint worse instead of better, which is exactly what motivated the few-sample reformulation at the beginning. Keeping one single shared embedding space where all grade are represented at the same time, and only imposing the ordinal structure at the final assignment step, allows every labeled example in the support set to help shape this same representation no matter which grade it belongs to, and this is a more sample-efficient use of the scarce labeled data than training a separate classifier for every grade pair. The second reformulation is about the temporal nature of psychological vulnerability. The general mapping from x to y in Section 2.1 implicitly treats the assessment as one single, self-contained event: a feature vector goes in, a risk grade comes out, and the problem is finished there. This kind of treatment cannot represent the phenomenon that motivated this research at the very beginning, that is, a student’s psychological condition evolves, sometimes gradually and sometimes suddenly, and the value of an assessment system partly lies in noticing this evolution before it turns into a crisis. So this paper reformulates the assessment problem, at the level of one individual student who is observed repeatedly over an academic term, as a sequential decision problem which unfolds over discrete assessment cycle indexed by t. At each cycle, an agent representing the assessment system observes a state that summarizes what is currently known about the student, chooses an action which represents an assessment decision such as keeping the current risk judgment, escalating it, or asking for additional observation, and then receives a reward that reflects how well this decision serves the two goals of accuracy and timeliness together. Framing the risk identification in this way replaces the single-shot question “what is this student’s risk level” with the repeated question “given everything observed so far, what is the best assessment decision right now,” and this is the structural basis for applying reinforcement learning to the problem, as well as the embodiment of the second innovation claimed in Section 1. A natural alternative for the sequential decision formulation would be simply repeating the static grading procedure of Section 3.1 (items 1–3) at every new assessment cycle, and treating the consecutive grades as an ordinary time series to be smoothed or extrapolated. This alternative is not chosen, because it treats every assessment cycle as an isolated event and it does not provide any principled mechanism for weighing the cost of an assessment action, such as the disruption caused by escalating a case or the resource cost of asking for additional observation, against the benefit that the information brings, and it also does not offer a natural way to reward the system for detecting an emerging risk trajectory before it turns into a crisis, instead of only reporting the current grade after it already happened. Treating the temporal problem as a sequential decision process with an explicit reward structure allows both of these considerations to be represented directly inside the model, instead of being handled as an ad hoc post-processing step applied on a sequence of otherwise independent static prediction, and it is exactly this property that Section 3.1 (items 4–7) makes use of when constructing the composite reward function. These two reformulations are not independent from each other; they are coupled by design, and this coupling is itself also part of the contribution of this paper. The few-sample hierarchical grading task produces, for a given student at a given point of time, both a discrete risk grade and a continuous embedding that locates the student inside the learned risk representation space. Instead of throwing away this information once a static grade has been produced, the reformulated sequential decision problem treats the static grade and its underlying embedding as part of the state that is observed by the reinforcement learning agent at the first assessment cycle, so the sequential module does not need to relearn the risk representation from raw behavioral data alone, but can instead build on the representation that is already learned under the few-sample paradigm. shows this coupling directly: the left panel depicts the few-sample task sampling structure that underlies the first reformulation, with several tasks each containing a small support set and query set of labeled example; the right panel depicts the sequential state-action-reward interaction loop that underlies the second reformulation, unfolding across the successive assessment cycle; and the bold arrow connecting the two panels represents the transfer of the static grade and embedding into the initial state of the sequential process. This coupled structure is exactly what Section 3 will formalize mathematically, first by establishing the few-sample grading model, then by establishing the sequential decision model, and finally by specifying how the two of them are joined together into one single trainable system. Mathematic formulation Model establishment Few-sample task structure and embedding representation Following the reformulation in Section 2.2, the pool of partially labeled student record is organized into a collection of episodic few-sample task, which are drawn from a task distribution p(T), and each task Tnis defined by a support set and a query set that are constructed over the same ordinal grade set {1,...,C}: where the superscript s and q denote the membership in the support and query subset respectively, K isthe number of labeled example per grade that is available in the support set (also called the “shot” count), C is the number of ordinal grade (also called the “way” count), and M is the number of query example whose label the model needs to infer within task Tn. A feature extraction network fθ(·), which is parameterized by θ, maps a raw aggregated feature vector x∈Rd into an embedding z=fθ(x)∈Rm, and for each grade c∈{1,...,C} inside task Tn. a class prototype is computed, following the prototype-based metric-learning paradigm for few-sample classification (), as the mean embedding of the support example that belong to that grade, so each task is represented, after only K labeled example per grade, by a small set of C prototype, instead of by a fully re-estimated classifier. Given a query example x, its grade is inferred through a distance-based softmax, which gives a higher probability to the grade whose prototype is closer to the query embedding in the learned space, and this is the mechanism through which the model can achieve few-sample generalization: the classification does not need re-training on new student, but only needs to re-compute the prototype from whichever small labeled set is available at the moment, as long as the embedding network fθitself has been trained to produce embedding that clusters meaningfully by grade, even when only a handful of example are seen. The squared Euclidean distance is used in () instead of a more elaborate learned metric, because once fθhas been trained under the episodic objective of (, ), the burden of capturing the relevant notion of similarity between student is already placed on the embedding network itself, and adding another learned distance function on top of the embedding would bring extra parameters that must also be estimated from the same scarce labeled data, which goes against the sampleefficiency goal that motivates the prototype-based design in the first place. It is also because of this choice that the grading model stays computationally light enough to be recomputed whenever a new support set becomes available, since the prototype computation in () only needs a forward pass through fθfollowed by a class-wise average, without any iterative optimization. Transfer-regularized meta-training objective The embedding network fθcannot be trained reliably from only the campus-specific labeled data, since this data is exactly the scarce resource that this paper is designed to work around. Instead, it is initialized from a broader pretrained representation θpre, which is obtained from related psychological or behavioral prediction task that has more abundant data, and it is then adapted to the campus assessment task through a combination of episodic meta-training and transfer-learning regularization. The source task that produces θpre is defined on the identical d-dimensional aggregated feature vector x as the target grading task, so no cross-domain feature alignment is required; what differs is only the label the network is trained to predict. Concretely, in the case studies of Section 4, where a calibrated simulation stands in for the inaccessible raw multi-cohort record (Section 4.1), the source task is a regression of the underlying continuous vulnerability trajectory itself, pooled across every meta-training-cohort student-week snapshot (7,259 snapshots drawn from the 109 cohort-A students, capped at the first 70 assessment weeks per student), trained for 200 full-batch Adam epochs at a learning rate of 10−2; this exercises the same feature-extraction network on a continuous, more abundantly available auxiliary signal before the network ever sees the sparsely labeled ordinal grade that the few-shot task of (-) uses. This construction keeps the source and target tasks on a shared feature schema while still separating what the pretraining stage is allowed to see (a continuous trajectory signal, available for every snapshot) from what the few-shot stage sees (the scarce discrete grade label), which is the property that () exploits. The meta-training loss for one single task penalizes the incorrect grade prediction on that task’s query set, and, because no single task can represent the full diversity of the student population, θ is optimized to perform well in expectation over the whole task distribution, instead of on any one particular task, where the quadratic penalty anchored at θpre implements the transfer-learning component of the first innovation: it discourages the embedding network from drifting too far away from its pretrained initialization so that it would overfit on the limited campus-specific task, while the regularization weight λreg controls how strongly the prior knowledge from the source domain is kept, relative to how much the network is allowed to specialize toward the target student population. Together, (, ) make up the training objective through which meta-learning and transfer learning are integrated into one single embedding-learning procedure, instead of being applied as two separate step that are only loosely connected. The regularization weight λreg works like a dial between two failure modes: if it is set too low, θ can drift far from θpre and risks reproducing the overfitting behavior that motivated the transfer-learning component at the first place, since the campus-specific episodic task alone is still a small-sample training signal, while if it is set too high, it will basically freeze the embedding at its pretrained state and stop the network from specializing on the behavioral and scale-score pattern that is specific to the target student population. The right operating point is determined empirically on held-out episodic task, as described in the model-solution discussion of Section 3.2, instead of being fixed beforehand. Ordinal hierarchical grade assignment Since the risk grade set is ordinal, not nominal, as established in Section 2.2, and following the threshold-based ordinal regression approach to graded classification and sorting problem (), the raw class-probability output of () is first converted into one single continuous vulnerability score that respects the ordering of the grade, before a discrete grade is finally assigned, which can be interpreted as the expected grade under the model’s own probability estimate, and therefore it changes smoothly as the underlying embedding moves between the prototype, instead of jumping suddenly like a plain arg-max classifier would do. The discrete hierarchical grade is then obtained by comparing this continuous score with a set of monotonically increasing threshold b0 γfp design choice argued for in Section 3, would operate near the high-ratio end, and the reward-weight ratio provides exactly the mechanism through which this institutional preference is encoded, without changing the underlying model structure. This lever is interpretable at the level of an institutional dial that an administrator can turn and observe the resulting Pareto trade-off, which is a weaker and more defensible claim than a feature-level explanation of any single decision; Section 5 discusses this distinction together with the explainable-machine-learning literature it should be read against. varies the discount factor γ ∈ {0.80,0.85,0.90,0.95,0.99} used for training the sequential policy (-)), and reports its effect on the cumulative reward, mean lead time, and falsealarm rate on the held-out student. The cumulative reward is maximized at the intermediate value (−114.1 at γ = 0.85 and −115.4 at γ = 0.90) and degrades a lot at the highest value tested (−149.6 at γ = 0.99), which indicates that weighting the distant future reward too heavily will destabilize the short-horizon policygradient update that the meta-reinforcement-learning procedure relies on, given that most student trajectory span well under 150 assessment cycle. The mean lead time increases only marginally and monotonically with γ (24.14 to 24.32 weeks), while the false-alarm rate rises a little bit (0.694 to 0.700), reflecting the slightly greater willingness of a more far-sighted policy to escalate early, in anticipation of a possible future crisis even when that crisis does not actually happen. If we put these results together, they indicate that the discount factor should be set in the moderately far-sighted range around γ = 0.85-0.90, instead of being pushed toward the myopic or the fully far-sighted extreme, so the sequential module’s anticipatory value can be balanced against the training stability under the realistically short student assessment horizon. Finally, reports the joint effect of the transfer-regularization weight λreg and the supportset size K on the grading macro-F1, through a heatmap spanning λreg ∈ {0,0.01,0.05,0.1,0.3} and K ∈ {5,10,20,30,50}, each cell averaged over 10 independently sampled evaluation seed (increased from 3 in an earlier version of this analysis) to make the reported trend less sensitive to individual-seed sampling noise; the standard deviation across these 10 seed ranges from 0.021 to 0.067 macro-F1 point depending on the cell, and is reported alongside the mean in the figure’s accompanying data table. At K = 5, increasing λreg from 0 to 0.3 raises macro-F1 from 0.813 to 0.824, a smaller but still directionally consistent improvement than in the earlier 3-seed estimate, and the largest improvement observed anywhere in the heatmap by row-wise column difference, while at K = 50 the same increase of λreg raises macro-F1 only from 0.857 to 0.867, a comparatively modest gain, and the intermediate column shows this same pattern getting weaker smoothly between these two extreme, with one exception: at K = 30 the highest regularization weight tested (λreg = 0.3) lowers macro-F1 slightly, from 0.846 at λreg = 0 to 0.832, a change of about 1.4 macro-F1 point against a standard deviation of 0.021-0.028 at that K, so it is a modest but not purely noise-level reversal. This is consistent with, rather than contradicting, the two-failure-mode description of λreg already given in Section 3.1: once the support set is large enough that episodic training alone shapes a reliable embedding, anchoring the network too strongly to the pretrained representation can start to cost more than it gives back, and K = 30 is exactly where this heatmap shows that crossover beginning, before K = 50’s larger support set again outweighs it. This interaction directly confirms the mechanism argued for in Section 3.1: the transfer-regularization term anchors the embedding network toward a pretrained representation, which is most valuable when the campus-specific episodic task alone gives too little signal to shape a reliable embedding by itself, and it becomes less and less necessary as the support set grows large enough for the episodic meta-training to succeed on its own. For an institution that wants to deploy this system, this indicates that the transfer-regularization weight should be set toward the higher end of the tested range whenever the anticipated labeled sample per grade is small, which is exactly the operating condition that motivated the transfer-learning component of the proposed model at the very beginning. Discussion Operational trade-offs under matched alert budgets This section interprets the case-study results reported in Section 4 against the operational constraint a counseling office actually faces, rather than reporting a further metric. The proposed meta-reinforcementlearning policy reaches its 24.3-week mean lead time at an escalation rate of 0.741 and an AUPRC of 0.514, while the static-reapplication baselines reach a substantially higher AUPRC (0.906-0.949 across the weekly-to-8-weekly screening frequencies of ) but only by triggering an alert under a strict High-or-Severe threshold and only a few weeks ahead of the crisis (3.3-7.1 weeks). Read in isolation, these two facts could be misread as the sequential module trading away most of its precision for its lead-time advantage; the budget-matched comparison of already shows that this trade is not simply “more alerts equal earlier alerts,” because at a matched false-alarm rate the proposed policy still anticipates the crisis earlier than the vanilla non-meta policy. What does not by itself resolve is the separate question this subsection addresses, namely whether the 24.3-week lead time is worth the 74.1% escalation rate in absolute institutional terms, given that the static baselines operate in an entirely different, much lower-triggering regime. The cost accounting of makes this concrete: over the held-out cohort’s trajectories, a weekly static-reapplication policy would require 91.8 screening events per student, while the full-composite-reward Meta-RL policy issues 63.0 escalate actions per student, a 31.4% reduction in counselor-facing flagged events relative to continuous weekly reassessment, and the accuracy-only-reward variant issues only 14.0 escalate actions per student, though it additionally issues 68.5 query actions per student, so its total count of resource-consuming action (escalate plus query, 82.5 per student) is in fact higher than the full-reward variant’s; only escalation specifically, not every resource-consuming action, corresponds to the counselor-facing flag that the screening-harm literature cautions against (). Read this way, escalation count specifically is the practically relevant quantity for alert fatigue, and on that quantity the full-composite-reward policy already represents a substantial reduction from continuous weekly reassessment, while the accuracy-only variant reduces it much further still, at the cost of a comparably sized increase in the number of lower-stakes query action, which request additional passive data collection rather than raising a flag to a counselor. Given the escalation-rate contrast above and the raw ablation numbers already reported in , the accuracy-only reward configuration (composite reward with w2 = w3 = 0) is best read as a compelling, practically preferable default for most deployment settings rather than as a stripped-down ablation of the proposed system. It gives up only 0.33 week of mean lead time relative to the full composite reward (23.99 vs. 24.32 weeks, negligible against the roughly 13-25-week standard deviation of individual student lead time already reported in ), while it lowers the escalation rate from 0.741 to 0.194 and raises the alert AUPRC from 0.514 to 0.737. This preference is reinforced, not weakened, by the regime-switching robustness check newly reported in : retraining both reward variant from scratch on a cohort whose crisis onset is an abrupt, discontinuous jump rather than the gradual mean-reverting accumulation of the main simulator reproduces the same qualitative pattern, a near-identical mean lead time between the two reward variant (28.27 vs. 28.22 weeks, a difference of under half a day, though nominally significant at p = 0.047 under a paired Wilcoxon test given the low variance of this metric in this cohort), together with a markedly lower escalation rate for the accuracy-only variant in this new cohort as well [0.463 vs. 0.000 for the full-reward variant, which here collapses to a policy that never actively escalates and instead relies entirely on the passive threshold-crossing behavior of the alert statistic ρi(t) in ()],. Contrary to the working hypothesis that motivated this robustness check, namely that the timeliness term’s contribution might become measurable once the risk trajectory departs from the gradual dynamics calibrated to the CES-documented statistics, no regime tested here, gradual or abrupt, produces evidence that the full composite reward delivers a materially longer anticipatory lead time than the accuracy-only reward. The honest reading of this evidence is that the composite reward’s timeliness term should not be presented as improving the achievable lead time, and the contribution statement and conclusion have been revised accordingly (Sections 1 and 6) to describe the timeliness and cost term as a tunable mechanism for shaping the escalation-versus-detection trade-off, exercised through the reward-weight-ratio sensitivity analysis of Section 4.3 (), rather than as a mechanism that by itself lengthens the anticipatory lead time beyond what an accuracy-only policy already achieves. For most deployment settings, and particularly for a counseling office where escalation volume is the binding operational constraint, this analysis recommends the accuracy-only reward configuration as the practical default, with the full composite reward retained as an available, explicitly higher-escalation operating point for an institution that has already exhausted its screening-harm tolerance elsewhere and specifically wants to bias the policy toward earlier, more frequent flagging. Marginal contribution of individual grading-model components Section 4.2 already reports, directly and not only in the sensitivity analysis, that the proposed grading model’s advantage over the plain prototypical network is small and not statistically significant at any tested support size, and that the attention-based matching-network baseline is statistically indistinguishable from the proposed model at K = 10 and significantly higher at K = 5. Taken together with the transferregularization sweep of , in which the benefit of the regularization weight λreg is concentrated at K ≤ 5 and becomes progressively smaller, and briefly reverses, as K grows toward 30-50, the picture that emerges is that episodic meta-training, not the specific prototype-based architecture or the pretrained-anchored regularization term individually, is the primary mechanism responsible for the large and consistently significant advantage over the non-episodic MLP and fine-tuned baselines documented in Section 4.2. This is consistent with, rather than contradicting, the first contribution as now stated in Section 1: the proposed grading model’s practical value lies in combining an existing episodic training paradigm () with a small, ultra-low-K-targeted regularization refinement and an ordinal thresholding mechanism, whose own ablation in Section 4.2 shows a smaller but still consistent contribution (macro-F1 reduces from 0.847 to 0.844 when removed), into one grading pipeline suited to the campus assessment setting, rather than in a claim that the prototype-based architecture itself is uniquely superior to every other episodically trained few-shot architecture. A more clearly separated architectural advantage, if one exists, would more plausibly come from a gradient-based inner-loop adaptation mechanism such as MAML-style meta-learning () than from a further variant of metric-based matching; the prototype-based design was favored in Section 3.2 specifically for its lower per-task computational cost, which stays a relevant practical consideration for a system intended to recompute a support set’s prototype on demand in an under-resourced counseling office, but this is a claim about deployment cost, not a claim that this specific architectural choice dominates every alternative few-shot method on accuracy. Relation to the few-shot learning and sequential clinical decision-making literature The macro-F1 advantage the proposed grading model achieves over a non-episodic classifier under scarce labeled data, more than twenty percentage point at K = 5, is directionally consistent with, though not directly comparable in magnitude to given the difference in task and modality, the gains reported for fewshot and prompt-based approach to psychological-text classification reviewed in Section 1 (–), and with the broader finding that few-shot learning narrows, without fully closing, the gap to a fully supervised model across biomedical time-series and behavioral-signal task (–); the present result extends this line of evidence to an ordinal, multi-source campus mental-health grading task, which none of the few-shot studies reviewed in Section 1 addressed directly. On the sequential side, the anticipatory lead time achieved by the proposed policy situates this work within the broader observation, already made for reinforcement-learning-based critical-care monitoring and knowledge-graph-based disease-progression modeling (, ), that a sequential decision formulation can surface an emerging risk trajectory earlier than a re-application of a static classifier, and the cost-accounting and budget-matched analysis of Sections 4.2 and 5.1 responds directly to a caution already present in that literature, namely that clinical adoption of reinforcement-learning-based monitoring remains at an early stage in part because its alerting behavior and resource cost are rarely quantified against the reassessment practice it would replace (). The finding that the composite reward’s timeliness term does not measurably lengthen the lead time relative to an accuracy-only reward, confirmed here under both the main gradual dynamics and the regime-switching robustness check, is also consistent with the general pattern in the adaptive and Bayesian sequentialmonitoring literature, in which a well-specified accuracy or discrepancy signal, updated as new data arrives, already carries most of a sequential system’s early-detection value without needing a separately engineered anticipation bonus (, ). Interpretability of the sequential policy The claims made in Sections 3.1 and 4.3 about the policy’s four-action space and the reward-weight ratio concern transparency at the level of the decision vocabulary and the institutional dial an administrator can turn, not feature-level explainability in the sense used in the explainable-machine-learning literature. Recent interpretable machine-learning work on psychiatric and youth-population risk prediction, for example posthoc explanation of the factor separating risk from protective influence in youth mood-state prediction during a public-health disruption (), interpretable prediction of clinically significant suicide risk among patients with major depressive disorder (), and feature-attribution-based modeling of self-reported suicidality in a clinical outpatient sample (), illustrates a materially stronger and more clinically established standard, namely an explanation of which specific input feature drove a given risk score or decision, typically through a post-hoc attribution method. No analysis in this paper substantiates a claim at that stronger standard, and the manuscript has been revised throughout (Sections 3.1 and 4.3) to state only the narrower claim the evidence actually supports, that the action space and reward-weight lever are semantically grounded and administratively tunable; a feature-level, post-hoc explanation of the grading model’s embedding or the sequential policy’s per-decision rationale, following the methodology of (–), is left as a direction for future work rather than a capability already delivered by the present system. Limits of a simulation-based validation Every quantitative result reported in Section 4 is obtained on a simulator whose free parameter are fit to the publicly documented scale, cohort structure, and modality composition of the CES/StudentLife study, but whose individual-level trajectory are not drawn from the study’s raw record, for the access-constraint reason stated in Section 4.1. The simulator’s generative mechanism (Equation () shares its structural form with the state representation the proposed sequential model is built around, so the case studies in Section 4 are better read as an internal-consistency and comparative-baseline demonstration, evidence that the proposed model outperforms the compared baseline when the environment matches the assumption the model was designed around, than as external validation that the reported effect size would transfer unchanged to a real deployment. The regime-switching robustness check of partially addresses this concern for the sequential module specifically, by showing that the qualitative pattern among the compared reward variant is not an artifact of the one particular dynamics calibrated to the CES statistics; it does not address the same concern for the static grading model, since both the main and the regimeswitching cohort share the same underlying feature-generation and grade-thresholding mechanism, nor does it substitute for validation against a real behavioral-sensing record with an independently obtained clinical label. Prospective validation on real campus data, ideally on a raw record from the CES/StudentLife study itself or a comparable publicly released longitudinal mobile-sensing mental-health dataset, therefore remains a necessary step before any deployment claim; the empirical claims in the abstract, Section 4, and the conclusion have been moderated throughout this revision to describe what a self-calibrated simulation study can demonstrate, internal consistency and comparative ranking among the compared method, rather than an estimate of real-world performance. Conclusion This paper has addressed the assessment of college students’ psychological vulnerability under a scarce, unevenly labeled data and an evolving risk trajectory that the static method cannot capture, using a calibrated simulation study whose evidentiary status is stated explicitly in Section 5.5 rather than treated as equivalent to validation on raw longitudinal record. It proposed a few-sample hierarchical grading model, which combines episodic meta-learning with a smaller transfer-regularization refinement and an ordinal thresholding mechanism, together with a reinforcement-learning-based sequential identification method that models the dynamic risk monitoring as a Markov decision process. Within this simulation study, the grading model outperformed the conventional deep-learning and threshold baseline by a large margin, especially under the smallest support size that reflects the real campus constraint, though its advantage over other episodically meta-trained few-shot architecture, the plain prototypical network and an attention-based matching network, was small and not consistently in its favor (Sections 4.2 and 5.2); episodic meta-training itself, rather than any one specific architectural choice on top of it, is the primary mechanism this evidence supports. The sequential module avoided the degenerate escalation behavior that a non-meta-adapted policy shows and achieved a markedly better alert precision-recall performance, but neither the main gradual-dynamics ablation nor a regime-switching robustness check with abrupt crisis onset () found evidence that its composite reward’s timeliness term delivers a measurably longer anticipatory lead time than a simpler accuracy-only reward; accordingly, the accuracy-only configuration, which reaches a comparable lead time at a much lower escalation rate, is recommended as the practical default, with the full composite reward available as a higher-escalation operating point (Section 5.1). The sensitivity analyses have also made clear how the institutional risk-tolerance parameter shapes the trade-off between early identification and intervention cost. Several limitation still motivate the future research. The case study relies on a calibrated simulation environment, instead of the raw longitudinal record, for the reason explained in Section 4.1, and this motivates a validation on real campus data under appropriate privacy safeguard, informed consent, and institutional data-governance procedure in the future (Ethics Statement); the regime-switching robustness check of Section 5.5 narrows but does not close this gap, since it still shares the main simulator’s featuregeneration mechanism. The feature representation also omits the qualitative peer- and counselor-reported information, which future work could try to incorporate through natural-language processing, so the sequential module’s state representation can be enriched. The reward weight that governs escalation and cost was explored only through a fixed sensitivity grid, instead of being elicited from institutional stakeholder, which suggests an extension toward preference-elicitation or multi-objective optimization. A feature-level, post-hoc explanation of the grading model’s and policy’s decisions, of the kind used in recent interpretable-machine-learning studies of psychiatric risk (Section 5.4), was also left outside the scope of the present system and is a natural next addition. Extending the sequential module to coordinate the assessment decision across multiple students under a shared counseling-resource constraint would also better reflect the real deployment condition. Funding Statement The author(s) declared that financial support was not received for this work and/or its publication. Footnotes Generative AI statement The author(s) declared that generative AI was used in the creation of this manuscript. The authors verify and take full responsibility for the use of generative AI in the preparation of this manuscript. Generative AI was used solely for language polishing and refinement, with no impact on the research content, data, or conclusions. Any alternative text (alt text) provided alongside figures in this article has been generated by Frontiers with the support of artificial intelligence and reasonable efforts have been made to ensure accuracy, including review by the authors wherever possible. If you identify any issues, please contact us. Publisher’s note All claims expressed in this article are solely those of the authors and do not necessarily represent those of their affiliated organizations, or those of the publisher, the editors and the reviewers. Any product that may be evaluated in this article, or claim that may be made by its manufacturer, is not guaranteed or endorsed by the publisher.