Video has become the dominant medium of online education, yet the learners who would gain most from remote learning are often the least able to receive it. A systematic review published in Discover Networks maps three decades of research on adaptive eLearning and adaptive video streaming and reaches a striking conclusion: not a single study in its corpus combines personalised learning, adaptive video delivery and developing-country deployment constraints in one architecture validated under African mobile-network conditions. The review, conducted to PRISMA 2020 standards by Mpako Enongene Ivo, Valery Nkemeni and Tanyi Emmanuel, screened 310 records and synthesised 39 studies published between 2013 and 2025, with Cameroon as the focal case.
The constraint profile the authors describe is compound rather than single-dimensional. Low-bandwidth environments are defined in the review as sustained throughput below roughly 2 megabits per second. On top of that sit mobile-data costs that exceed the United Nations Broadband Commission’s affordability target: in 2024, the median price of the entry-level 2 GB monthly mobile-broadband basket in Africa stood at 4.2 percent of gross national income per capita, more than double the 2 percent target. Learners typically access platforms through entry-level smartphones with 1 to 2 GB of RAM, no dedicated graphics processor and software decoding only, while electricity supply itself can be unreliable. These factors interact multiplicatively, producing a deployment environment qualitatively different from the broadband-saturated conditions for which most eLearning platforms were designed.
The review organises the literature into three pillars. Pillar A covers adaptive learning: systems that use learner modelling, machine learning and data-driven personalisation to adjust content, sequencing and pace to the individual. Pillar B covers adaptive video delivery, principally adaptive bit-rate streaming protocols such as MPEG-DASH and HTTP Live Streaming, in which content is encoded at multiple quality levels and a client-side controller selects a representation for each short segment based on observed network and buffer state. Pillar C covers developing-country deployment, including infrastructure, cost and adoption constraints. Each pillar is individually well populated, but the intersections are thin, and the central three-way intersection is empty.
The quantification is stark. Across the 39-study corpus, 39 percent of studies address adaptive learning alone, 23 percent adaptive streaming alone and 28 percent developing-country deployment alone. Only three studies sit at the intersection of adaptive learning and adaptive streaming, and just one of them, the DQAMLearn framework developed within the European Horizon 2020 NEWTON project by Moldovan and Muntean, integrates the pillars substantively. That framework groups mobile devices into classes, maintains per-class multimedia profiles, and introduces two mechanisms, VQAMap and QuTCompute, that map objective video-quality metrics to subjective Mean Opinion Score values and compute the quality thresholds at which controlled reductions can preserve a predefined user-perceived quality. It was validated in a subjective study with sixty participants across six categories of educational clips, reporting a mean quality-of-experience rating of 79.19 on a 0–100 acceptability scale. But its network and device assumptions were calibrated for European deployment, and no study in the corpus is validated under African mobile-network conditions.
The technical mismatch runs deep. Deep reinforcement learning has become prominent in recent adaptive bit-rate research, but the algorithms are trained on networks running at 3 to 5 megabits per second and above, and offer no meaningful adaptation below 2 megabits. At 500 kilobits per second, the review notes, they pin the lowest quality tier and remain there for the duration of the session. Commercial deployments of DASH and HLS typically begin their encoding ladders at 360p, around 400 to 700 kilobits per second, which is often unsustainable at typical African throughput. The authors argue the mismatch is structural rather than parametric: algorithms trained on 10 to 50 megabit networks have simply never encountered the decision space that a 500-kilobit link presents.
The corpus does contain partial antecedents. A grey-literature undergraduate thesis from Ashesi University in Ghana, the CC4LB algorithm, targets the 150 kilobits to 2 megabits regime characteristic of African 2G and 3G access, prioritising connection consistency over peak quality and accepting a lower mean bit rate in exchange for reduced rebuffering; the review treats it as a relevant antecedent while flagging its limited evidentiary weight. Bin Waheed and colleagues characterised how segment duration, swept from zero to twenty seconds, trades adaptive responsiveness against coding efficiency in an HLS architecture with a content delivery network validated on 4G/LTE. Venkatesh Naik and Madhavi proposed a recurrent neural network with long short-term memory for bit-rate selection in collaborative eLearning, showing that temporal context can outperform stateless heuristics, though the review declines to rely on the specific magnitudes because the published units are inconsistent. On the encoding side, Siddique and colleagues applied temporal masking to motion vectors and residual frames in lecture video, achieving competitive SSIM and PSNR quality at reduced bit rate.
On the learning side, the picture is one of pedagogical sophistication paired with a systematic blind spot. A systematic review of sixty-three articles by Gligorea and colleagues documents the dominance of the closed-loop architectural pattern, in which systems collect learner data, infer learner state, select content, deliver feedback and iterate, alongside persistent challenges of data privacy, system complexity and integration with mainstream learning management systems. Morze and colleagues demonstrated that meaningful adaptation of materials, sequencing and pace is achievable within standard Moodle through conditional activities and lesson branching, lowering the implementation barrier for resource-constrained institutions. Yet across this body of work, the review observes, network bandwidth and video delivery are treated almost universally as content properties rather than as first-class adaptation dimensions.
The developing-country evidence base reinforces the scale of the problem. A review of Kenyan universities reported that 90.8 percent of e-learners expressed dissatisfaction with e-learning delivery, that 2,000 sub-locations have below 50 percent 3G coverage, and that broadband services cover only 17 percent of the country’s landmass. Studies of learning management system adoption at major African universities found extremely low active use, with fewer than 60 active users at Makerere University and fewer than 10 at the University of Nairobi. In Cameroon specifically, the review finds no peer-reviewed study addressing eLearning at the technical platform level; the country ranks among the bottom ten African economies for mobile-broadband subscriptions at roughly thirty per hundred inhabitants against a regional average of fifty-two, and the closest peer-reviewed technical antecedent is Osee and colleagues’ hybrid quality-of-experience prediction work on Cameroonian 3G and 4G networks, which fuses network-side measurements with user-satisfaction scores.
From this gap analysis the authors derive eight design requirements: mobile-first operation at a minimum viable throughput of 150 kilobits per second; support for entry-level devices with 1 to 2 GB of RAM and software decoding only; bandwidth-efficient encoding; user-visible control over the quality-cost trade-off with per-session data budgets; offline capability with pre-downloaded lessons and opportunistic synchronisation; lightweight adaptation logic that degrades gracefully when learner-model updates are delayed; a cloud-elastic hybrid backend; and adoption-aware design treating individual, organisational, environmental and technological factors as first-class inputs. These feed a preliminary four-layer reference architecture combining a hybrid-cloud backend, an adaptation engine integrating learner modelling with a conservative low-bandwidth bit-rate controller, a progressive-web-app client capped at 480p software-decoded playback, and a context-awareness module that monitors signal, battery, storage and usage patterns to defer large synchronisations until charging or Wi-Fi.
The authors are explicit that the architecture is a literature-grounded design argument, not a validated system, and that its planned evaluation at the University of Buea, with approximately sixty participants across Cameroonian operators and throttled 2G, 3G and 4G conditions, remains future work. The review also acknowledges its own limits: single-reviewer screening with supervisory spot-checking rather than independent dual screening, possible under-representation of Francophone-African research, publication bias against negative deployment results, and a corpus skewed toward the COVID-era publication surge. The central message, however, stands on the quantified evidence: the components of an adaptive, low-bandwidth eLearning platform exist and have been validated in isolation, but the integrative engineering and the contextual validation under African conditions have not yet been done. Until they are, the review concludes, the promise of adaptive eLearning will remain inaccessible to the learners who stand to benefit most.
Subject of Research: A systematic literature review of adaptive eLearning and low-bandwidth video delivery for developing-country higher education
Article Title: Adaptive eLearning and video delivery for low-bandwidth environments: a systematic literature review
Article References: Ivo, M. E., Nkemeni, V., & Emmanuel, T. (2026). Adaptive eLearning and video delivery for low-bandwidth environments: a systematic literature review. Discover Networks, 2(1), Article 36. https://doi.org/10.1007/s44354-026-00050-1
Image Credits: AI Generated
DOI: 10.1007/s44354-026-00050-1
Keywords: adaptive learning, adaptive bitrate streaming, eLearning, low-bandwidth networks, sub-Saharan Africa, Cameroon, systematic review, PRISMA, quality of experience, digital divide, video on demand, higher education
News Source: Denise Maddox. (October 29, 2026). Systematic review finds no eLearning system built for Africa’s slow networks. Scienmag.
