In the absence of microscopy and rapid tests, malaria diagnosis is presumptive in many clinics in Africa, leading to overuse of drugs and masking other causes of fever. It has been proposed, based on pre-test data (e.g., symptoms, clinical notes, patient voice), to use standard ML models as triage tools. We searched PubMed and Scopus for studies from 2010 to 2026 that trained and evaluated standard ML models on pre-test clinical data to paint a high RQ1 for malaria diagnosis or on voice or clinical text for other infectious diseases (RQ1) and followed PRISMA 2020. We included 146 studies (RQ1=43; RQ2=103) out of 2598 studies. A modified PROBAST assessed the risk of bias. No malaria studies used speech, and only four used free text or language models. Excluding laboratory reference standard, 27 (63%) out of 43 studies for malaria employed 95% or greater accuracy, found in 14 out of 17 studies. Those with laboratory labelling (and independent adjudications) found discrimination modest, with AUCs between 0.60–0.95. However, chance-level predictions were observed when new speakers, recording devices, and entirely new datasets were used; only 14 of 103 studies mention independent data. These data justify pre-test malaria triage only if subsequent models include parasitological classification, patient-level speaker-independent splits, symptom-matched controls, external validation, and combined and separate reporting from each input. Keywords: clinical decision support, malaria, machine learning, natural language processing, speech analysis, systematic review